MTBF Model for AVs - From Perception Errors to Vehicle-Level Failures (IV2022)

Summary

  • from Intel + Mobileye
  • RSS modelはPerception errorを考慮していなかったため、perception system error と vehicle-level failures (= collision) を紐付けた評価が必要だった
  • Perception errorを考慮した自動運転車両の安全性のEvaluation modelとして、Mean Time Between Failure (MTBF) models の提案
    • Situation
    • Vehicle mission profile (speed range, road characteristics, traffic density, time of day)
    • Potentially dangerous Situation
    • Perception Errorの種類
    • の重み和を用いて衝突するprobabilityを計算する、というのが基本的な考え方

Background

  • “Responsibility Sensitive Safety” (RSS) model
    • Planning error における safety frameworks
    • Perception errorはしない仮定を置いていた

Method

Safety Relevant and Severe Perception Errors

  • Typeを2つに分ける

    • Type.1 False Positive / Underestimate distance or speed
      • 実際には車がいないのに、認識器はいると判断
      • 要らない急ブレーキかけちゃう危険性
    • Type.2 False Negative / Overestimate distance or speed.
      • 実際には車がいるのに、認識器はいないと判断
      • 前方のブレーキに間に合わない
  • Safety-Relevant Perception Error (KPI)

    • ISO 26262 defines four classes of severity, S0 to S3
    • Safety decision of the AV planning module (RSS) is changed

Potentially Dangerous Traffic Situation

  • Sever Perception Error = S2 - S3 Severity
  • Hazardous (Dangerous) Traffic Situation
    • Typeごとに危険因子を分解

Statistical Model

  • Estimate Hazardous Traffic Situation Rate
    • Use naturalistic driving data (e.g. HighD, REM, etc.)
    • Obtain information on how long/how often a vehicle performs certain actions (decelerating, accelerating, lane changes, driving with certain speed, etc.)
  • $$ MTBF ^ {-1} = \sum_{m=\text { Mission }} p_{m} \sum_{i=\text { SpeedRange }} p_{i, m}\left[\sum_{t=\text { ErrorType }} \lambda_{p_{t, m, i}} \times p_{S_{t, m, i}}\right] $$
    • $ p_{m} $: the occurrence probability of mission profile $m$
    • $ p_{i, m} $: the occurrence probability of the speed range $ i $ (e.g. the range from 100 km/h to 130 km/h) for mission profile $ m $
    • $ \lambda_{p_{t, m, i}} $: perception error rate
      • hardwareそのものの故障とsoftwareの能力不足両方含んでいる
    • $ p_{S_{t, m, i}} $: the probability of being within a relevant traffic situation

  • 今回はi: speed range としたが、自由に設定できる
    • 危険なsituationも自由に設定可能

Model Input

  • perception error rates を得るために解析を行う
    • $ d _ {per} > d _ {RSS} > d _ {real} $ -> Type 2 Error

Experiment

HighD dataset

  • https://www.highd-dataset.com/
  • Type II perception errorsのみの解析
  • データセット(における視野角)的に1sec以上程度しか車が存在しないため、Type I errorは解析しても微妙なため
  • lostの方が影響が大きいため、velocity and distance errors も無視

model

  • $ p_{S_{i}} $ のmodelを以下の通りにする(situationごとに定義できる)
    • $$ p _ {S _ {i}}=p _ {d _ {i}}+p _ {a _ {i}} \times p _ {a _ {T T C, i}}+\left(1-p _ {a _ {i}}-p _ {d _ {i}}\right) \times p _ {c _ {T T C, i}} $$
    • We define “close” to the lead vehicle, as a TTC5 (time-to-collision) of less than 5 seconds.
    • 1項
      • $ p _ {d _ {i}} $: the probability of a lead car decelerating in speed range i
    • 2項
      • $ p _ {a _ {i}} $: the probability that the lead car is sufficiently close
      • $ p _ {a _ {T T C, i}} $: the probability that the lead car accelerates
    • 3項
      • $ p _ {c _ {T T C, i}} $: the probability that the lead car is sufficiently close
      • $ p _ {c _ {i}} = 1-p _ {a _ {i}}-p _ {d _ {i}} $: the probability that the lead car drives with constant speed
  • 解析
    • 80-100 km/h, 100-130 km/h and 130- 180 km/h の区分で分けることに

Extracting Perception Quality Requirements

  • $$ \lambda=\sum_{i} \lambda_{i} \times p_{i} \times p_{S_{i}}=\hat{\lambda} \times \sum_{i} p_{i} \times p_{S_{i}}=: \hat{\lambda} \times \kappa $$
    • $ \hat{\lambda} $: the perception error rate
  • 最終的なmodel

From Perception Errors to Vehicle-Level Failures

  • Lyft dataset で LiDAR detectionのみで perception errorを算出
  • the severe perception miss rate (frames with misses per second) is 17/5040 s = 12.1 errors/second
    • 速度は考慮無し
  • the overall vehicle-level MTBF is 1523 second (0.4 hours) で目標にまだ届いていない

Discussion

  • Datasetに評価するべきscenarioが包括されているか
  • 評価にData augmentation を適応する時は、biasが入りやすいので慎重に
  • error duration の解釈
    • 1 miss * 500 frameと100miss * 5 frameでは解釈が異なるはず