Glossary

Probability & statistics

Probability確率

A number from 0 to 1 (0% to 100%) expressing how likely an event is.→ 1.1

Independence独立

One event happening does not change how likely the other is. For independent events, the probability that both happen is found by multiplication.→ 1.1, 1.2

Mutually exclusive排反

Two events that cannot happen at the same time. For mutually exclusive events, the probability that either happens is found by addition.→ 1.1

Conditional probability条件付き確率

The probability that B happens given that A has happened, written P(B | A) — “the share of B within the world where A has happened”.→ 1.2

Bayes’ theoremベイズの定理

A formula that flips the direction of a conditional probability, from P(A | B) to P(B | A). It lets you calculate a probability after seeing evidence, such as an alarm.→ 1.2

Prior / posterior probability事前確率・事後確率

The prior is your belief before seeing evidence; the posterior is the belief updated after seeing it. For rare events, a low prior leads to many false alarms.→ 1.2

Likelihood ratio尤度比

The ratio of the probability of a piece of evidence (such as a passed test) if the system is safe to its probability if the system is unsafe. Multiplying the prior odds by it gives the posterior odds. A test that the unsafe case also passes has a likelihood ratio near 1, so passing it is weak evidence.→ 1.2, 5.10

Probability distribution確率分布

A list of the possible outcomes and how likely each one is.→ 1.3

Binomial distribution二項分布

The distribution of the number of successes in n trials with success probability p — for example, the number of detections in 10 tests of a detector with a 90% detection rate.→ 1.3

Poisson distributionポアソン分布

The distribution of how many times a rare event occurs within a given distance or time. Its shape depends only on the average count m.→ 1.3, 1.8

Normal distribution正規分布

A symmetric, bell-shaped distribution centred on the mean. About 68% lies within ±σ of the mean, about 95% within ±2σ and about 99.7% within ±3σ.→ 1.3

Expected value期待値

The sum of each possible value multiplied by its probability — an average weighted by probability.→ 1.4

Exposure遭遇頻度

The share of driving spent in a given condition (rain, night, congestion and so on). The overall accident rate is the expected value of the per-condition rates weighted by exposure. Also used in functional-safety risk assessment (the E in ASIL).→ 1.4, 5.6

Variance / standard deviation分散・標準偏差

The variance is the average of the squared deviations from the mean; the standard deviation σ is its square root, expressing the size of the spread in the original units.→ 1.4

Law of large numbers / central limit theorem大数の法則・中心極限定理

The more trials you run, the closer the average gets to the true value (law of large numbers). The average of many independent values approaches a normal distribution even when the original distribution is not normal (central limit theorem).→ 1.5

Standard error標準誤差

The size of the spread of an estimated average (or proportion): the spread of a single result σ divided by √n, so quadrupling the number of trials halves it.→ 1.5, 1.6

Monte Carlo methodモンテカルロ法

Estimating probabilities or averages by running many random trials (simulations). Precision improves only with the square root of the number of trials.→ 1.5, 5.7

Confidence interval / confidence level信頼区間・信頼度

A range, estimated from data, in which the true value is likely to lie. A 95% confidence interval means that if you repeated the same estimation many times, about 95% of the intervals would contain the true value. For safety requirements of the form “no more than …”, the one-sided upper bound (estimate + 1.64 × standard error for 95%) is used.→ 1.6

Point estimate / interval estimate点推定・区間推定

Estimating the true value from data as a single number is a point estimate (e.g. a detection rate of 90%); estimating it as a range is an interval estimate (e.g. 86–94%).→ 1.6

Sample variance (dividing by n − 1)標本分散

The variance used to estimate the true spread from sample data. Dividing the sum of squared deviations by n − 1 instead of n avoids underestimating the spread on average.→ 1.4, 1.6

t-distributiont分布

A distribution used to build confidence intervals for a mean when there are few data. Its tails are wider than those of the normal distribution, so intervals widen as the data shrink (with 6 data points, about 2.57 replaces the “2”).→ 1.6, 1.7

Bootstrapブートストラップ

A method for finding the scatter (confidence interval) of an estimate by repeatedly resampling the data with replacement and redoing the estimation. It works for complicated calculations and quantiles, but only reflects the random scatter of the data, not errors in the model’s assumptions.→ 1.9

Hypothesis testing仮説検定

A method for deciding whether you can claim a difference, by checking how unlikely the observed result would be if there were no difference. The assumption of “no difference” is called the null hypothesis.→ 1.7

p-value / significance levelp値・有意水準

The p-value is the probability, assuming there is no difference, of a result at least as extreme as the one observed. If it is below a significance level fixed in advance (usually 5%), the result is called “statistically significant”. A p-value is not the probability that there is no difference.→ 1.7

Type I error / type II error / power第1種の誤り・第2種の誤り・検出力

Claiming a difference when there is none is a type I error (probability α); missing a difference that really exists is a type II error (probability β). The probability 1 − β of correctly finding a real difference is the power, which rises as the number of tests grows.→ 1.7

Multiple comparisons多重比較

Making many comparisons at once. Even when there is no difference at all, a 5% significance level produces a chance “significant” result about once in twenty comparisons, so countermeasures are needed, such as a stricter threshold (the Bonferroni correction divides α by the number of comparisons).→ 1.7

Rule of three3の法則

After n trials (or n km) with zero failures, the upper bound on the failure rate is about 3/n at 95% confidence. To show a target of “no more than one accident per X km”, you need about 3X km without an accident.→ 1.8

Heavy-tailed distribution / extreme value裾の重い分布・極値

A distribution in which large outliers occur more often than in a normal distribution. How to handle rare, extreme values matters a great deal in safety evaluation.→ 1.3, 1.9

Return level / return period再現レベル・再現期間

The value exceeded only once in n trials (or n years) is the return level, and n is the return period — the same idea as a “once-in-a-century flood”. The more trials, the worse the worst value you meet.→ 1.9

Extreme value theory / peaks over threshold極値理論・しきい値超過法

A method that collects only events beyond a threshold (such as near misses), fits the shape of their tail and extrapolates how likely rarer events (such as accidents) are. With a strict enough threshold, the excesses approach a generalized Pareto distribution whatever the original distribution. Results depend heavily on the assumed tail shape.→ 1.9

Machine learning & AI

Machine learning機械学習

A way of having a computer learn rules (a model) from data instead of having people write the rules.→ 2.1

Supervised / unsupervised / reinforcement learning教師あり学習・教師なし学習・強化学習

Supervised learning learns the relationship between inputs and answers from labeled data; unsupervised learning finds groupings or unusual items in data without answers; reinforcement learning learns, by trial and error, actions that earn a larger reward.→ 2.1

Regression / classification回帰・分類

Within supervised learning, regression predicts a number (such as a position 3 seconds ahead) and classification predicts a category (such as pedestrian or cyclist).→ 2.1, 2.2, 2.4

Model / parameter / inferenceモデル・パラメータ・推論

A model is the computation that turns an input into a prediction; parameters are the adjustable knobs that set its behavior. Training fixes the parameters, and using the resulting model to answer new inputs is called inference.→ 2.1

Generalization汎化（汎化性能）

The ability to predict correctly on new data that was not used for training. It is measured on data held back from training.→ 2.1, 2.6

Loss function損失関数

A single number measuring the gap between predictions and correct answers. Learning means adjusting the parameters to make this number smaller.→ 2.2

Residual残差

The difference between the correct value and the prediction (y − ŷ) — the “prediction gap” for each data point.→ 2.2

Mean squared error / RMSE / MAE平均二乗誤差・RMSE・MAE

The mean of the squared residuals is the MSE; its square root, the RMSE, gives the typical gap in the original units; the mean of the absolute residuals is the MAE. Because the MSE weights large gaps heavily, it is easily pulled by outliers.→ 2.2

Least squares最小二乗法

A method that sets the parameters of a line (or another model) so that the sum of squared residuals is smallest. For linear regression, the answer can be computed directly with a formula.→ 2.2

Extrapolation外挿

Predicting outside the range covered by the data. A model that fits well inside the range can be far off outside it.→ 1.9, 2.2

Gradient descent勾配降下法

A method that computes the slope (gradient) of the loss function and nudges the parameters downhill, step by step, to find where the loss is smallest.→ 0.2, 2.3

Learning rate学習率

The number (η) that sets how far gradient descent moves in each step. Too small and learning is slow; too large and it overshoots, swinging back and forth or diverging.→ 2.3

Stochastic gradient descent / mini-batch / epoch確率的勾配降下法・ミニバッチ・エポック

Gradient descent that estimates the slope from a small, randomly chosen subset of the data (a mini-batch) at each step. One pass through the whole dataset is called an epoch.→ 2.3

Local minimum局所解

A point where the loss is lowest in its neighborhood but not necessarily lowest overall. In neural networks, where training ends up depends on the initial values and on randomness.→ 2.3

Logistic regression / sigmoid functionロジスティック回帰・シグモイド関数

A classification method that turns the straight-line expression z = wx + b into a probability between 0 and 1 with the sigmoid function 1/(1 + e^−z). Despite “regression” in its name, it is used for classification.→ 2.4

Oddsオッズ

How many times more likely an event is to happen than not: p/(1 − p). In logistic regression, the logarithm of the odds is a straight line in the inputs.→ 2.4

Cross-entropy / maximum likelihood交差エントロピー・最尤推定

A loss function widely used for classification. It grows as the probability given to the correct answer falls, and a confident mistake is penalized very heavily. Minimizing cross-entropy is the same as maximizing the probability of the observed data (the likelihood) — maximum likelihood estimation.→ 2.4

Decision thresholdしきい値（判定の）

The cut-off that turns a model’s probability into a decision such as “detect / do not detect”. Lowering it reduces misses and increases false detections; raising it does the opposite.→ 2.4, 2.7

Softmax functionソフトマックス関数

A function that turns the scores for three or more classes into probabilities that are at least 0 and add up to 1. Used in the final stage of image-recognition neural networks.→ 2.4

Decision tree / Gini impurity決定木・ジニ不純度

A model that decides by stacking yes/no questions in the shape of a tree. Each question is chosen to reduce the mixing after the split (the Gini impurity, 1 − Σp²) the most. A small tree can be read by people as a set of rules.→ 2.5

Ensemble (random forest / gradient boosting)アンサンブル（ランダムフォレスト・勾配ブースティング）

Predicting with the majority vote or average of several models. The benefit is larger the more independent the models’ mistakes are, and smaller between models trained on the same data.→ 2.5

Overfitting過学習

Fitting the training data so closely that performance drops on new data. Prevented with techniques such as regularization.→ 2.6

Underfitting学習不足

A model so simple that its error is large on both the training data and new data. Improved with a more expressive model or more useful inputs.→ 2.6

Training / validation / test data訓練データ・検証データ・テストデータ

Training data are used to learn the parameters, validation data to choose settings such as model complexity, and test data once, at the end, to measure the true performance.→ 2.6, 2.8

Cross-validation交差検証

Repeating training and validation with different splits of the data and averaging the results. Reduces the uncertainty of the evaluation when data are scarce.→ 2.6

Regularization (ridge regression)正則化（リッジ回帰）

Adding a penalty on the size of the weights (such as λ × the sum of squared weights) to the loss, giving a smoother model that works on new data. Too strong a penalty leads to underfitting.→ 2.6

Early stopping早期終了

Stopping training when the error on the validation data starts to rise, to prevent overfitting.→ 2.3, 2.6

Precision / recall / miss rate適合率・再現率・見逃し率

Recall is the share of real objects that were detected; precision is the share of detections that were real. Neither one alone tells you how good a detector is. The miss rate (false negative rate, FNR) is 1 − recall. Recall is also called sensitivity or the true positive rate (TPR), and precision the positive predictive value (PPV). For the false positive rate (FPR), 1 − FPR is called the specificity (true negative rate, TNR).→ 1.2, 2.7

False positive / false negative誤検知・未検出

A false positive reports something that is not there; a false negative misses something that is. In automated driving, both unnecessary hard braking and missed detections can lead to hazards.→ 1.2, 2.7

Confusion matrix / accuracy混同行列・正解率

A table counting results in four cells: correct detections (TP), misses (FN), false alarms (FP) and correct rejections (TN). Accuracy — the share of all cases classified correctly — tends to look high when the target is a minority, so it cannot judge performance on its own.→ 2.7

F1 scoreF1スコア

A single number combining precision P and recall R: 2PR/(P + R). It does not grow when only one of the two is high.→ 2.7

ROC curve / AUCROC曲線・AUC

The curve of false positive rate (horizontal) against recall (vertical) as the threshold moves, and the area under it. AUC is the probability that, picking one real and one false case at random, the real one gets the higher score; 0.5 means random guessing.→ 2.7

IoU (intersection over union)IoU（重なり率）

In object detection, the area where the predicted box and the true box overlap, divided by the area they cover together. A detection counts as correct when the overlap is above a set value (for example 0.5).→ 2.7, 3.4

Distribution shift分布シフト

A mismatch between the data used for training and the data encountered in real use. A major cause of AI performance degradation.→ 2.8

Covariate / prior / concept shift共変量シフト・事前確率シフト・コンセプトシフト

The three types of distribution shift. Covariate shift is a change in the inputs (snowy regions, night-time, a replacement camera); prior shift is a change in the share of targets (how many pedestrians there are); concept shift is a change in the relationship between inputs and answers (new kinds of vehicles, rule changes).→ 2.8

Data leakageデータの漏れ

Test scores coming out higher than a model’s real ability, for example because similar data from the same trip, day or place appear in both the training and the test data. Prevented by splitting data by groups such as trips, days or regions.→ 2.6, 2.8

Out-of-distribution detection分布外検出

Methods for noticing that an unfamiliar input, far from the training data, has arrived — used as a trigger to behave on the safe side, for example by restricting the function.→ 2.8, 3.7

Neural networkニューラルネットワーク

A computation built from many layers of dot products and simple transformations — the core model behind today’s AI, including image recognition.→ 0.2, 3.1

Neuron / activation functionニューロン・活性化関数

The building block of a neural network. It adds a bias to the weighted sum of its inputs and passes the result through an activation function (such as the sigmoid function or ReLU). The activation function “bends” the calculation, which is what gives stacking layers its point.→ 3.1

ReLUReLU

An activation function: max(0, z) — 0 when z is negative, z itself when it is positive. Cheap to compute and widely used in deep networks.→ 3.1

Deep learning / hidden layerディープラーニング（深層学習）・隠れ層

Learning with neural networks that stack many layers between input and output (hidden layers). Earlier layers come to represent simple features, later layers more complex things.→ 3.1

Backpropagation / chain rule誤差逆伝播法・連鎖律

A method that passes the output error back layer by layer, obtaining the slopes of all the weights in a single backward pass. It relies on the chain rule: slopes (rates of change) travel as products of the rates along the way.→ 3.2

Vanishing / exploding gradients勾配消失・勾配爆発

In deep networks, the chain-rule product gets long, so the slopes reaching the early layers become far too small (vanishing) or too large (exploding). Prevented with ReLU, careful initialization, normalization and residual connections.→ 3.2

Automatic differentiation自動微分

Software that works out slopes (derivatives) automatically and exactly from the steps of a calculation — what makes backpropagation usable in real development.→ 3.2

CNN / convolution / filterCNN（畳み込みニューラルネットワーク）・畳み込み・フィルタ

A neural network that slides a small window (a filter) across the whole image and “multiplies and adds” over the covered region (convolution). Using the same weights at every position keeps parameters few and finds the same feature wherever it appears. The resulting array is called a feature map.→ 3.3

Poolingプーリング

Summarizing a feature map block by block (for example, keeping only the largest value in each 2×2 block). Reduces computation and adds tolerance to small shifts of objects.→ 3.3

Shortcut learning近道学習

Learning to decide from cues that merely happen to be in the data (background, texture, capture conditions) instead of the intended ones (such as an object’s shape). If the test data share the same bias, accuracy alone will not reveal it.→ 3.3, 3.9

Object detection / bounding box物体検出・バウンディングボックス

The task of answering, for each object in an image, a box giving its position and size (the bounding box), its category and a confidence. Box error (regression) and category (classification) are learned together.→ 3.4

Semantic / instance segmentationセグメンテーション（セマンティック・インスタンス）

Semantic segmentation answers a category (road, sidewalk, car and so on) for every pixel; instance segmentation also separates individual objects of the same category.→ 3.4

NMS (non-maximum suppression)NMS（非最大値抑制）

A procedure for removing duplicate boxes for the same object: repeatedly keep the most confident box and remove boxes that overlap it heavily (IoU above a threshold). When separate objects overlap in the image, it can remove one of them.→ 3.4

AP / mAP (average precision)AP・mAP（平均適合率）

The standard evaluation metric for object detection. AP is the area under the precision–recall curve as the threshold moves; mAP is the mean AP across categories. Note that it counts every miss with the same weight.→ 3.4

Danger-weighted metric危険度で重みづけた指標

A metric that counts errors with weights for how dangerous they are (distance to the road user, time to collision and so on). Among models with the same overall recall, one that misses more nearby road users scores much lower. The weights are fixed in advance, with their grounds.→ 3.10

Consecutive misses続けて見逃す確率

The probability of missing the same road user in several consecutive images. Because the causes of misses (glare, occlusion and so on) persist, it is much larger than the per-image miss rate multiplied together. It is evaluated on video by counting how long runs of misses last.→ 3.10

Paired comparison (McNemar’s test)対応のある比較（マクネマー検定）

A way to compare two models on the same test data, judging whether the difference is due to chance from the counts b and c of cases only one of them got right. If z = (b − c) / √(b + c) exceeds about 2, there is a difference. It can tell smaller differences apart than comparing on separate data.→ 1.7, 3.10

Test oracleテストオラクル

The means of judging whether a test result is correct (the correct answer or a judging criterion). For AI perception, being unable to prepare a correct answer for every input is the central problem of testing.→ 3.11

Metamorphic testingメタモルフィックテスト

Testing that checks relations of the form “if the input is changed like this, the output should change (or stay the same) like that”. Flipping images or changing their brightness allows many tests without labels. A relation being kept does not mean the answer is correct, so it supplements tests with correct answers.→ 3.11

Negative flipネガティブフリップ

A case that the previous model got right but the updated model gets wrong. It happens even when average performance improves, so regression testing checks the number and content of such cases.→ 3.11

Slice testingスライス試験

Testing that splits the test data by condition (night, rain, long range and so on) and examines performance per condition to find weak spots. Telling weakness apart needs enough cases per condition.→ 3.11, 5.9

Attention / query, key, value注意機構（アテンション）・クエリ・キー・バリュー

A mechanism that decides, with weights, how much to draw on each part of the input. The dot products of what is being looked for (the query) with each element’s features (the keys) are turned into weights by the softmax function, and the contents (the values) are averaged with those weights. The weights are not necessarily the reasons for a decision.→ 3.5

TransformerTransformer

A neural network that stacks many layers built around attention. It can relate distant elements directly and is used widely, for text, images, scene prediction and more; its computation grows with the square of the number of tokens.→ 3.5, 3.6

Tokenトークン

The unit of input a Transformer handles: words or parts of words in text, small patches in an image, individual vehicles and pedestrians in a driving scene, and so on.→ 3.5

Foundation model / pre-training / fine-tuning基盤モデル・事前学習・微調整

A huge model trained on masses of unlabeled data (pre-training) and then reused for many purposes through extra training on a little data (fine-tuning) or through written instructions.→ 3.6

LLM / VLM (large language model / vision-language model)LLM・VLM（大規模言語モデル・視覚言語モデル）

A foundation model that handles text is an LLM; one that handles both images and text is a VLM. An LLM writes by predicting the probability of the next token from the sequence so far.→ 3.6

Temperature (in generation)温度（生成の）

The number the scores are divided by before the softmax function when generative AI picks the next token. A high value spreads out the answers and makes unlikely ones more frequent; a low value gives almost the same answer every time.→ 3.6

Hallucinationハルシネーション

Generative AI producing untrue content that sounds plausible, in a confident tone. It happens because the model learns “is this a plausible continuation?”, not “is this true?”.→ 3.6

Data contaminationデータの混入（ベンチマークの汚染）

When the questions and answers used for evaluation end up in a model’s training data. If the model has “memorized” the answers, its accuracy comes out higher than its real ability on unseen questions. Check with unpublished or newly written questions.→ 3.6

AI-as-judge scoringAI による採点

Having another AI score the quality of a generative AI’s answers. It is fast and cheap, but biased — for example towards longer answers or answers shown first — so it is compared with human scoring before being relied on.→ 3.6

Aleatoric / epistemic uncertainty不確実性：偶然的（アレアトリック）・認識的（エピステミック）

Aleatoric uncertainty (from the Latin for “dice”) comes from randomness or ambiguity in the data itself and does not shrink with more data; epistemic uncertainty (from the Greek for “knowledge”) comes from a lack of knowledge or data and can be reduced by collecting more data.→ 3.7

Calibrationキャリブレーション

Whether the confidence an AI reports (for example, 90%) matches how often it is actually right. If they match, the model is said to be well calibrated.→ 3.7

Reliability diagram / ECE (expected calibration error)信頼性図・ECE（期待較正誤差）

A reliability diagram groups answers by confidence and compares each group with its actual accuracy; points below the diagonal mean overconfidence. ECE is the count-weighted average gap between confidence and accuracy.→ 3.7

Temperature scaling温度スケーリング

A calibration method that divides the scores by a temperature T before they become probabilities, correcting how confidence is reported. T is chosen on validation data. Which answer is most probable does not change, so accuracy is unchanged.→ 3.7

Robustness / corruption頑健性（ロバスト性）・劣化

The property of a decision not changing when the input changes a little. It covers natural corruptions such as rain, dirt, backlighting and blur, and deliberate small changes (adversarial examples); the two are evaluated separately.→ 3.8

Adversarial example敵対的サンプル

An input with changes too small for people to notice that makes an AI badly misjudge. Used to evaluate robustness.→ 3.8

Adversarial training / data augmentation敵対的訓練・データ拡張

Adversarial training creates adversarial examples during training and learns to answer them correctly; data augmentation trains on images with added rain, brightness changes and the like. Adversarial training often lowers accuracy on unchanged inputs a little.→ 3.8

Explainability (XAI)説明可能性（XAI）

Techniques that examine which parts of the input an AI’s decisions rely on and present this in a form people can understand. Explanations can be local (a single decision) or global (the model’s overall tendencies). A plausible explanation is not necessarily a correct one.→ 3.9

Saliency map / occlusion顕著性マップ・遮蔽

A saliency map pictures how fast the output changes when each pixel changes a little (the slope with respect to the input). Occlusion hides part of the image and measures how much the confidence drops.→ 3.9

SHAP / feature attributionSHAP・特徴の寄与

A method that fairly divides the difference in output from an average situation among the inputs, showing how much each input pushed the answer up or down. For a straight-line model it equals weight × (input − average).→ 3.9

Counterfactual explanation反実仮想の説明

An explanation of the form “had the input been different in this way, the answer would have changed like this”, showing the smallest change needed to change the answer.→ 3.9

Automated driving

Levels of driving automation自動運転レベル

The six levels, 0 to 5, defined in SAE J3016. They are distinguished by whether the human or the system performs the driving and whether the system is limited to certain conditions.→ 4.1

Operational Design Domain (ODD)運行設計領域

The range of conditions an automated driving system is designed to work in — road types, weather, speed, time of day and so on. The same performance can mean different things in different ODDs.→ 1.2, 4.1

ODD boundaryODD の境界

The limit of a condition included in the ODD (for example, up to what rainfall). It is placed before the point where performance drops (the performance boundary), with a margin for the time the fallback takes and the error of detection, and defined in quantities the vehicle can measure while driving. The grounds, including the trade-off between exposure and availability, are recorded.→ 4.1

Dynamic driving task (DDT)動的運転タスク（DDT）

The real-time work of driving: operating acceleration, braking and steering, and noticing and responding to other vehicles, pedestrians and so on (monitoring and responding). Choosing destinations and routes is not included.→ 4.1

Takeover request / DDT fallback引き継ぎ要求・フォールバック

Fallback is the response when the system has a fault or is about to leave its ODD. At Level 3, the system asks the person to take over the driving (a takeover request) and the person does so; from Level 4 up, the system handles it itself.→ 4.1

Minimal risk condition (MRC)最小リスク状態（MRC）

The state a vehicle is brought into to keep the risk of a crash as low as possible when driving cannot continue — for example, stopping on the shoulder or in the lane with the hazard lights on.→ 4.1

Perception / planning / control認知・判断・操作

The basic flow of automated driving: recognize the surroundings (perception), decide how to drive (planning), and operate the steering and brakes (control).→ 4.2

Latency遅れ（レイテンシ）

The time from when the sensors capture a scene until the vehicle actually starts to respond. The processing times of the stages add up, and because the vehicle keeps moving during the delay, the stopping distance grows.→ 4.2

Object tracking追跡（トラッキング）

Linking the objects detected in each frame to the same objects in earlier frames and following how their position and speed change. It can make up for a miss in a single frame and gives speed estimates.→ 3.4, 4.2

Drive-by-wireバイワイヤ

Operating the brakes and steering by electrical signals instead of mechanical linkages. It is used in the control stage of automated driving, and redundancy against faults becomes important.→ 4.2

Sensor fusionセンサフュージョン

Combining information from several sensors, such as cameras, LiDAR and radar, so that they compensate for each other’s weaknesses.→ 4.3

LiDARライダー

A sensor that measures the distance to surrounding objects from the time it takes laser light to bounce back.→ 1.3, 4.3

Radarレーダー

A sensor that sends out radio waves and measures distance from the time until they bounce back, and relative speed from the shift in their frequency (the Doppler effect). It copes well with rain, fog and darkness but is poor at telling shapes apart.→ 4.3

Point cloud点群

A set of 3D points obtained from LiDAR and similar sensors. Each point has a position (and a reflection strength).→ 3.4, 4.3

Stereo camera / disparityステレオカメラ・視差

Two cameras side by side. Distance is found from how far the same object is shifted between the left and right images (the disparity). The further away, the smaller the disparity, and the distance error grows with the square of the distance.→ 4.3

Kalman filterカルマンフィルタ

A method that, at every time step, combines a prediction and a new measurement, weighting each by the size of its error (its variance). Widely used for tracking, sensor fusion and localization.→ 4.2, 4.3

Localization自己位置推定

Estimating where the vehicle is on the map and which way it is facing, by combining GNSS, dead reckoning and map matching.→ 4.4

HD map高精度地図

A map recording lane shapes, stop lines, signs and so on to an accuracy of a few to a dozen or so centimetres. It gives information beyond sensor range, but can disagree with reality, for example because of roadworks.→ 4.4

Dead reckoning推測航法（デッドレコニング）

Finding position by adding up the motion measured from wheel rotation and the IMU. It works without outside references, but its error builds up with the distance travelled.→ 4.3, 4.4

Map matching地図照合

Finding the vehicle’s position by comparing LiDAR point clouds or lane markings seen by cameras with the map and searching for the best overlap. Similar-looking places can be confused.→ 4.4

GNSS / RTKGNSS・RTK

GNSS is the general name for systems that find position from satellite signals (GPS and others). The error is usually a few metres, but a few centimetres with RTK, which uses corrections from base stations. It is unavailable or much less accurate in tunnels and urban canyons.→ 4.3, 4.4

Motion prediction周りの動きの予測

Estimating where surrounding vehicles and pedestrians will be a few seconds from now. Rather than a single answer, it is often given as several futures with probabilities, such as “waits 70%, crosses 30%” (multimodal prediction).→ 4.5

Route, behaviour and trajectory planning経路計画・軌道計画

Deciding the roads to the destination (route planning), what to do next (behaviour decision) and the position and speed for the next few seconds (trajectory planning). Candidates are evaluated, the best is chosen, and the plan is redone in a short cycle.→ 4.5

Cost functionコスト関数

An expression of how bad a candidate way of driving is, as a weighted sum of danger, comfort, progress and so on. It is the same idea as a loss function in machine learning, and the choice of weights shapes the behaviour.→ 2.2, 4.5

Safe-distance model (RSS)安全な車間距離のモデル（RSS）

An approach that sets by formula quantities such as the following distance that lets the vehicle stop, given its response time and braking, even if the car ahead brakes hard, and then keeps to them. It is used as a “safety envelope” that checks the planner’s output from outside.→ 4.5

End-to-end drivingEnd-to-End 型

An approach in which a single large AI model learns directly from inputs such as camera images to driving actions. Contrasted with the modular approach, which builds each part separately.→ 4.6

Imitation learning / reinforcement learning模倣学習・強化学習

Imitation learning copies human driving as a model to follow; reinforcement learning learns by trial and error guided by rewards. In imitation learning, errors tend to compound once the model enters “drifted” states absent from the demonstrations.→ 4.6

Open-loop / closed-loop evaluation開ループ評価・閉ループ評価

Open-loop evaluation measures, scene by scene in recorded data, the gap between the model’s output and the demonstration. Closed-loop evaluation tests the model where its decisions move the vehicle and the results become the next inputs (as in simulation). Good open-loop scores do not guarantee good closed-loop driving.→ 4.6

Autonomous emergency braking (AEB)衝突被害軽減ブレーキ

A function that detects the risk of colliding with a vehicle or pedestrian ahead and brakes automatically. False activations (unnecessary hard braking) can also lead to hazards.→ 1.2

Time to collision (TTC)衝突余裕時間

The time until a collision if things continue at the current speeds. The shorter it is, the more dangerous the situation; it is used as a near-miss indicator.→ 1.8

Disengagement / driver interventionドライバー介入

During automated-driving tests, a safety driver taking over control for safety. The number of disengagements is one of the metrics of a test.→ 1.3

SOTIF & safety

Functional safety (ISO 26262)機能安全

The approach, and the standard, for reducing hazards caused by faults (malfunctions) of electrical and electronic systems to an acceptable level.→ 4.7

ASILASIL（自動車用安全度水準）

Automotive Safety Integrity Level: in functional safety, the safety level set for each hazardous event a fault could cause, from its severity (S), exposure (E) and controllability (C). There are four levels, A to D, with D the strictest; events that reach none of them are QM (ordinary quality management).→ 4.7, 5.4

Failure rate / FIT故障率・FIT

The failure rate is the probability that a part fails per hour. Because it is so small, it is often expressed in FIT (failures in time), the number of failures per billion hours (1 FIT = 10⁻⁹ per hour).→ 4.7

Fail-safe / fail-operationalフェイルセーフ・フェイルオペレーショナル

Fail-safe means moving to a less dangerous state when a fault occurs. Fail-operational means keeping the necessary functions despite a fault and continuing until the vehicle has stopped safely; it is needed at Level 3 and above, where a person cannot take over at once.→ 4.1, 4.7

Cybersecurity (ISO/SAE 21434) / TARAサイバーセキュリティ（ISO/SAE 21434）・TARA

The approach, and the standard, for protecting vehicles from malicious attacks. The procedure of identifying threats and rating risk from the feasibility and impact of attacks is called TARA (threat analysis and risk assessment).→ 4.4, 4.7

Common cause共通原因

A single cause — dense fog, heavy rain, a power-supply fault — that weakens several components or sensors at once. Estimating the benefit of redundancy by multiplication while overlooking common causes tends to overestimate safety.→ 1.1, 2.5, 4.7

β factorβ ファクター

A fraction expressing the size of common causes: the share β of one component’s failures that come from causes which make the other component fail at the same time. The probability that both fail is then about βp + {(1 − β)p}², which departs from the independent p² even when β is only a few per cent.→ 1.1

SOTIF (ISO 21448)意図した機能の安全性

Safety of the intended functionality: the approach, and the standard, for reducing hazards that arise even without a fault — from functional insufficiencies (limits of performance or specification) or reasonably foreseeable misuse.→ 5.1

Functional insufficiency機能不十分

An umbrella term for insufficiencies of specification (situations that were not anticipated) and performance insufficiencies (sensing or recognition that is not good enough). Distinguished from faults.→ 5.1, 5.3

Triggering conditionトリガー条件

A situation that brings out a functional insufficiency and initiates hazardous behaviour — for example, sun glare, heavy rain or an unfamiliar object.→ 5.1, 5.3

Pairwise testing (combinatorial testing)ペアワイズ試験（組み合わせ試験）

Instead of trying every combination, choosing tests so that every combination of two conditions is tried at least once. It makes defects and triggering conditions that appear in combination easier to find with few tests.→ 5.3, 5.6

Design of experiments実験計画法

A way of choosing which tests to run: decide the parameters (factors) and their steps (levels), then investigate the effect of each factor and of their combinations efficiently with few tests. Orthogonal arrays and pairwise testing are tools of it.→ 5.6

Orthogonal array直交表

A list of tests built so that, for any two columns, every combination of levels appears the same number of times. For example, the 81 combinations of four factors at three levels are covered with 9 tests (L9).→ 5.6

Interaction交互作用

When the effect of one condition depends on the level of another. For example, there is no problem at night alone or in rain alone, but performance drops on a rainy night.→ 5.3, 5.6

The four areas4つの領域

SOTIF divides situations into four areas by “known or unknown” and “hazardous or not”, and works to shrink both the known hazardous and the unknown hazardous areas.→ 5.2

Good–Turing estimateグッド・チューリングの推定

A method that estimates “the probability that the next event is of a kind never seen before” as N₁/N, where N₁ is the number of kinds that appeared exactly once among N recorded events. It assumes events keep occurring under the same conditions.→ 1.8, 5.2

Reasonably foreseeable misuse合理的に予見可能な誤使用

Use that differs from the designer’s intent but is readily predictable from human behaviour. SOTIF treats it as one source of hazards.→ 5.1

Hazardous behaviour危険な振る舞い

Behaviour of the vehicle that can lead to danger, resulting from a functional insufficiency brought out by a triggering condition — not stopping when it should, braking hard when it should not, and so on.→ 5.1

Residual risk残留リスク

The danger that remains after measures have been taken. It cannot be made zero, so it is shown to be below the acceptance criteria.→ 5.1, 5.5, appendix

Hazard / harmハザード・危害

Harm is physical injury or damage to health; a hazard is a potential source of that harm.→ 5.4

HARA (hazard analysis and risk assessment)ハザード分析とリスク評価

A procedure for identifying possible hazardous events and assessing their risk from factors such as severity, exposure and controllability.→ 1.4, 5.4

STPAシステム理論に基づくハザード分析

An analysis method that models a system as a control structure — who commands what and how each part responds — and identifies unsafe control actions that can lead to hazards.→ 5.4

Unsafe control action (UCA)危ない制御動作（UCA）

In STPA, a control action that leads to danger in a particular situation. UCAs are identified by four types: not providing it causes a hazard, providing it causes a hazard, wrong timing or order causes a hazard, and wrong duration causes a hazard.→ 5.4

FTA / FMEAFTA・FMEA（故障の木解析・故障モード影響解析）

Fault tree analysis puts an undesired event at the top and breaks its causes down with AND and OR, estimating combinations of causes and their probability. Failure mode and effects analysis lists each way a part can fail and traces its effect on the whole from the bottom up.→ 5.4

Acceptance criteria受入れ基準

Criteria stating how safe is safe enough — for example, a fatality rate lower than that of human drivers.→ 1.8, 5.5

Validation target妥当性確認目標

The amount or value of evidence needed to argue that the acceptance criteria are met — for example, the distance that must be driven without an accident.→ 1.8, 5.5

Verification / validation検証・妥当性確認

Verification checks “did we build it to the specification?”; validation checks “is it safe enough for its real purpose?”.→ 5.5

ALARPALARP（合理的に実行可能な限り低く）

“As low as reasonably practicable”: the principle of weighing the effort of measures against their effect and taking every reasonably practicable measure, to bring risk as low as possible.→ 5.5

Positive risk balanceポジティブ・リスク・バランス

The idea that introducing automated driving reduces harm overall compared with human driving; one of the approaches to setting acceptance criteria for automated driving.→ 5.5

Surrogate measure代わりの指標（サロゲート指標）

Because accidents themselves are rare and hard to count, safety is estimated from more frequent events that can lead to accidents, such as near misses with a short time to collision (TTC). The assumed ratio to accidents must be checked separately.→ 1.8, 5.5

Scenario-based testingシナリオベース検証

Testing by systematically selecting situations (scenarios) likely to lead to hazards and varying their conditions.→ 5.6

Functional, logical and concrete scenarios機能・論理・具体シナリオ

Three levels of abstraction for scenarios: a functional scenario is an outline in words, a logical scenario gives the parameters describing the situation and their ranges, and a concrete scenario fills in specific values to make a single test.→ 5.6

Naturalistic driving data自然運転データ

Data recorded from everyday driving as it happens. It shows how often situations occur and how their parameters are distributed.→ 5.6

XiL (MIL, SIL, HIL, VIL)X-in-the-loop

An umbrella term for simulation-based testing that combines models, software, hardware and real vehicles step by step.→ 5.7

Gap between simulation and realityシミュレーションと現実の差

The difference between a simulation result and the result of trying the same situation in reality, arising from imperfect models of vehicle motion, sensors and the behaviour of others. Its size is known only within the range actually compared.→ 5.7

Multi-pillar assessment複数の柱による評価

The approach of assessing the safety of automated driving by combining methods of different kinds, such as audits of documentation, simulation, test tracks and public-road testing.→ 5.7

Importance sampling重要度サンプリング

A method that deliberately tests rare, dangerous situations more often and corrects the results with weights, to estimate their overall likelihood efficiently.→ 1.8, 5.8

Falsification反例探索（ファルシフィケーション）

A method that actively searches, by optimization and similar means, for situations that break a safety requirement (counterexamples). Counterexamples found are dealt with as known hazards. Finding none is not proof of safety.→ 5.8

Robustness value安全の余裕の指標（ロバストネス）

A number expressing by how much a requirement was met, such as “minimum gap − required gap”. Below 0 the requirement is broken; falsification searches for situations that make this value small.→ 5.8

ISO/PAS 8800ISO/PAS 8800（道路車両の安全と AI）

“Road vehicles — Safety and artificial intelligence” (2024): a publicly available specification setting out how to use AI components safely. It describes deriving requirements for AI components and datasets from system requirements, verifying them and monitoring in operation, supplementing functional safety and SOTIF.→ 5.9

Dataset qualityデータセットの品質

Aspects of whether data used for training, evaluation and testing can serve as safety evidence: coverage, correctness (few label errors), independence, representativeness, traceability and so on.→ 2.8, 5.9

Label errorsラベルの誤り

Errors in the labels given as ground truth (omissions, mix-ups, misaligned boxes). Because the difficult examples people tend to miss are also those AI tends to get wrong, label errors in evaluation data can make apparent performance look better than it really is.→ 5.9

Safety caseセーフティケース

A structured argument, built from evidence and reasoning, that a system is safe enough. Often drawn as a GSN diagram.→ 5.10

GSNGSN（ゴール構造表記法）

Goal Structuring Notation: a way of drawing safety cases. Claims (goals) are rectangles, argument strategies parallelograms, evidence circles and context rounded rectangles, arranged so that each claim above is supported by the claims and evidence below it.→ 5.10

Defeater反論（ディフィーター）

A doubt of the form “in what cases would this claim fall?”. Listing defeaters and eliminating them with evidence strengthens an argument; those that cannot be eliminated are written down openly.→ 5.10

Safety performance indicator (SPI)安全性能指標（SPI）

An indicator decided in advance to watch safety after release, such as the frequency of accidents, near misses or departures from the ODD. When it crosses a set line, the cause is investigated. It is a means of checking that the claims and context of the safety case still hold.→ 5.10, 5.11

Leading and lagging indicators先行指標と遅行指標

Lagging indicators count outcomes, such as accidents; leading indicators count things that happen short of an accident, such as near misses or hard braking. Leading indicators are more numerous, so changes are noticed sooner.→ 5.11

Control chart管理図

A chart of counts or measurements in time order, with lines marking the range of usual variation (for example three standard deviations from the mean). Crossing a line suggests a cause beyond the usual and is investigated. The method comes from quality control in manufacturing.→ 5.11

OTA updateOTA（無線でのソフトウェア更新）

Over-the-air update: updating a vehicle’s software through a network connection without bringing it to a factory or dealer. It allows problems to be fixed quickly, but regression testing and update management are essential so that updates do not introduce new errors.→ 5.11

Further reading

• How to read symbols: Notation, 0.2 Reading math notation
• To report an error or suggest a term to add, please use the contact form