Critical Analysis of Prototype Autonomous Vehicle Crash Rates

Name: Critical Analysis of Prototype Autonomous Vehicle Crash Rates | Six Scientific Studies from 2015-2018
Brand: SAE International
Price: 245.99 EUR
Availability: OnlineOnly

Six Scientific Studies from 2015-2018

Richard Young(Author)

SAE International (Publisher)

1st Edition

Published on 30. November 2021

249 pages

E-Book

PDF with Adobe-DRM

System requirements

978-1-4686-0342-2 (ISBN)

€245.99incl. 7% vat

System requirements

for PDF with Adobe-DRM

E-Book Single Licence

Available for download

Description

More details

Other editions

Content

Cover
Title Page
Copyright Page
Table of Contents
Table of Figures
Table of Tables
Foreword
Preface
Author's Note Regarding Abbreviations, Keywords, and Phrases
Acknowledgments
Introduction
CHAPTER 1 Methods
1.1 IRR Estimation Method
1.2 IRR Confidence Limit Estimation Method
1.3 Heterogeneity Best Practice Violation
1.4 Matching AV Crashes to Police Records
1.5 Reason for Not Using Statistical Significance Tests
References
CHAPTER 2 The First Major AV Safety Study: Schoettle and Sivak (2015)
2.1 Summary of Schoettle and Sivak (2015)
2.1.1 Original Schoettle and Sivak (2015) Data: AV versus CV Crash Rates
2.1.2 Unrecorded Crash Selection Bias: CV Crashes (Table 2.1 , Note f)
2.1.2.1 Adjustment for Unrecorded CV Crashes
2.1.3 AV Site Selection Biases: Environment, City Streets (Table 2.2 , Note d)
2.2 Critical Analysis of Schoettle and Sivak (2015) Study
2.2.1 Minor Systematic Error: Wrong Percentage of Unrecorded CV Injury Crashes (Table 2.2 , Note f)
2.2.1.1 Correction of Minor Systematic Error: CV Crashes (Table 2.2 , Note f)
2.2.2 Potential Site Selection Bias from Non-CA AV Tests (Table 2.2 , Note c)
2.2.2.1 Removal of Possible Site Selection Bias from Non-CA AV Tests
2.2.3 Best Practice Violation: Combining Heterogeneous Companies (Table 2.2 , Notes b and c)
2.2.3.1 Remove Best Practice Violation: Stratify Heterogeneous Companies
2.2.4 Best Practice Violation: Combining Heterogeneous Severities (Tables 2.1 and 2.2 , Notes b and f)
2.2.4.1 Remove Best Practice Violation: Stratify Heterogeneous Crash Severities
2.3 Reanalysis of Corrected Schoettle and Sivak (2015) Data
2.4 New Methods to Remove Unrecorded Crash Selection Bias and Site Selection Bias
2.4.1 First New Method to Remove Unrecorded Crash Selection Bias: Use Only Police-Recorded AV Crashes
2.4.2 Second New Method to Remove Unrecorded Crash Selection Bias and AV Site Selection Bias: AV Trend Analysis
2.4.2.1 Trend Analysis of Google AV Cumulative Crash Rate
2.4.2.2 Hypotheses for Upward Trend in 2015 Google AV Cumulative Crash Rate
2.4.2.2.1 Hypothesis 1. Changes in Google Testing Protocols.
2.4.2.2.2 Hypothesis 2. Negative Learning by the Google Neural Network.
2.4.2.2.3 Hypothesis 3. Upward Trend in CV Cumulative Crash Rate.
2.4.2.2.4 Hypothesis 4. Downward Trend in Test Drivers' Cumulative Disengagement Rate.
2.4.2.3 Tests of Hypotheses 3 and 4 for Upward Trend in 2015 Google AV Cumulative Crash Rate
2.5 Summary of Reanalysis of Schoettle and Sivak (2015) Data
References
CHAPTER 3 Google AV Crash Data versus Naturalistic Crash Data: Blanco et al. (2016)
3.1 Summary of Blanco et al. (2016): Google AV versus SHRP2 CV
3.1.1 Google AV Original Crash Data
3.1.2 SHRP2 CV Original Crash Data
3.1.3 Google AV versus SHRP2 CV Crash Rates
3.2 Epidemiologic Analysis: Blanco et al. (2016) Google AV versus SHRP2
3.2.1 Blanco et al. (2016) Replicated: Google AVs Reduce Crash Rate
3.2.2 Blanco et al. (2016) versus Schoettle and Sivak (2015)
3.3 Critical Analysis of Blanco et al. (2016): Site Selection Bias
3.3.1 Site Selection Bias Does Not Explain Why Google AVs Reduce the Severe Crash Rate (Table 3.1A)
3.3.2 Site Selection Bias Explains Why Google AVs Appear to Reduce the Major PDO Crash Rate (Table 3.1B)
3.3.3 Site Selection Bias Explains Why Google AVs Appear to Reduce the Minor PDO Crash Rate (Table 3.1C)
3.4 Standardizing the AV Crash Data Adjusts for Site Selection Bias
3.4.1 Standardize Google AV Crashes to SHRP2 Data Sites
3.4.2 Summary and Discussion of Standardized Results
3.4.3 Limitations of Standardization
3.5 Blanco et al. (2016): Google AV versus U.S. National CV Comparison is Confounded by Unrecorded Crashes
3.5.1. SHRP2 versus U.S. National
3.5.2 Discussion of Blanco et al. (2016) Unrecorded Crash Results
References
CHAPTER 4 Crash Selection Bias and Site Selection Bias: Dixit et al. (2016)
4.1 Summary of Dixit et al. (2016)
4.2 Critical Analysis of Dixit et al. (2016)
4.2.1 Injury-Only CV Crash Systematic Error: (Table 4.1 , Note f)
4.2.2 (Table 4.1 , Notes b and f)
4.2.3 Improving Accuracy: Statewide CV MVMT (Table 4.1 , Note g)
4.2.4 Partly-Corrected Dixit et al. (2016) Data (Table 4.2)
4.2.5 Unrecorded Crash Selection Bias: (Table 4.2 , Notes b and f)
4.2.6 AV and CV Crashes (Table 4.2 , Notes b and f)
4.2.7 Discussion and Conclusion of Dixit et al. (2016) Review
References and Endnotes
CHAPTER 5 Crash Selection Biases and Site Selection Biases: Teoh and Kidd (2017)
5.1 Phase 1 "Mountain View": 2009-December 31, 2015
5.1.1 Teoh and Kidd's (2017) Phase 1 Original Data
5.1.1.1 AV Crashes
5.1.1.2 AV Mileage
5.1.1.3 CV Crash-Involvement
Three CV Crash-Involvement Selection Criteria for Mountain View, CA
5.1.1.4 CV Mileage
CV Mileage Selection Criteria for Mountain View, CA
5.1.1.5 Summary of Teoh and Kidd's (2017) Phase 1 Original Data
5.1.2 Critical Analysis of Teoh and Kidd (2017): Phase 1
5.1.2.1 AV Crash Selection Bias 1: "Police-Reportable" AV Crashes (Table 5.1 , Note b)
Adjustment Method: Use Police-Recorded AV Crashes.
5.1.2.2 AV Crash Selection Bias 2: Censoring Transition Crashes (Table 5.1 , Note b)
Correction Method: Count Transition Crashes as AV Crashes.
5.1.2.3 AV MVMT Site Selection Bias 1: Inclusion of Non-CA Mileage (Table 5.1 , Note c)
Correction Method: Remove Non-CA Mileage.
5.1.2.4 AV MVMT Site Selection Bias 2: Unspecified AV Mileage in Other Cities (Table 5.1 , Note c)
Correction Method: Not Possible.
5.1.2.5 CV Crash Selection Bias 1: Multi-Vehicle CV Crashes (Table 5.1 , Note f)
Correction Method: Count Multi-Vehicle CV Crashes as One Crash.
5.1.2.6 CV Crash Selection Bias 2: Highway CV Crashes (Table 5.1 , Note f)
Correction Method: Remove CV Crashes on Highways.
5.1.2.7 CV Crash Selection Bias 3: Counting Passenger Vehicles Only (Table 5.1 , Note f)
Correction Method: Count All CV Types.
5.1.2.8 CV MVMT Minor Update for 2015 (Table 5.1 , Note g)
Correction Method: Use Actual 2015 Data.
5.1.2.9 CV MVMT Site Selection Bias (Table 5.1 , Note g)
Correction Method: None Available.
5.1.2.10 CV MVMT City Street Site Selection Bias (Table 5.1 , Note g)
Correction Method: None Needed After Removal of Highway CV Crashes.
5.1.3 Teoh and Kidd's (2017) Partly-Corrected Phase 1 Data and AV IRR
Why Only Partly-Corrected?: AV Crash Uncertainty Remains
5.1.4 Discussion and Conclusion to Teoh and Kidd's (2017) Phase 1
5.2 Phase 2 "All States": January 1, 2016, to August 31, 2016
5.2.1 Summary of Teoh and Kidd's (2017) Phase 2
5.2.2 Critical Analysis of Teoh and Kidd's (2017) Phase 2
5.3 Trends in Teoh and Kidd's (2017) Data
5.3.1 The Trend in AV Cumulative Crash Rate
5.3.2 The Trend in CV Cumulative Crash Rate
5.3.3 The Effect of the CV Cumulative Crash Rate on the AV Cumulative Crash Rate
5.3.4 The Trend in the Test Drivers' Cumulative Disengagement Rate
5.3.5 The Effect of the Test Drivers' Cumulative Disengagement Rate on the AV Cumulative Crash Rate
5.3.6 The Trend in the ADS Cumulative Disengagement Rate
5.3.7 The Effect of the ADS Cumulative Disengagement Rate on the AV Cumulative Crash Rate
5.4 Discussion of Reanalyzed Teoh and Kidd (2017) Data
References and Endnotes
CHAPTER 6 Heterogeneity and Crash Selection Bias: Favarò et al. (2017)
6.1 Summary of Favarò et al. (2017) Original Data
6.1.1 Original Data in Tabular Format
6.1.2 Original Total AV and CV Data in Epidemiologic Format
6.2 Critical Analysis of Favarò et al. (2017)
6.2.1 Best Practice Violation 1: Heterogeneous Vehicle Types (Table 6.2 , Note h)
6.2.2 Best Practice Violation 2: Heterogeneous CV Sites (Table 6.2 , Notes b and f)
6.2.3 Unrecorded Crash Selection Bias (Table 6.2 , Note f)
6.3 Discussion of Favarò et al. (2017)
6.3.1 Corrections Not Made in Favarò et al.'s (2017) Data
6.3.2 Favarò et al. (2018) Study
References and Endnotes
CHAPTER 7 Crash Selection Bias and Site Selection Bias: Banerjee et al. (2018)
7.1 Summary of Banerjee et al. (2018)
Banerjee et al.'s (2018) Method to Estimate Median APM
7.2 Critical Analysis of Banerjee et al. (2018)
7.2.1 Minor Error in Nissan Median APM
7.2.2 Data Anomalies
7.2.3 Failure to Replicate Median DPM
7.2.4 Terminology Definition Questions
7.2.4.1 Ambiguity in DPM Definition
7.2.4.2 Ambiguity in DPA Definition
7.2.4.3 Failure to Replicate Median APM per Manufacturer
7.2.5 AV Crash Rates and IRRs Estimated Using Standard Methods
7.2.6 Unrecorded Crash Selection Bias
7.2.7 Site Selection Bias
7.2.8 Conclusion to Analysis of the Banerjee et al. (2018) Study
References
CHAPTER 8 Overall Summary of AV IRR Estimates and Issues
8.1 Overview
8.2 Overall Summary of AV IRR Estimates
Chapter 2: Schoettle and Sivak (2015) [8.1]
Chapter 3: Blanco et al. (2016) [8.2]
Chapter 4: Dixit et al. (2016) [8.3]
Chapter 5: Teoh and Kidd (2017) [8.4]
Chapter 6: Favarò et al. (2017) [8.5]
Chapter 7: Banerjee et al. (2018) [8.6]
Final IRR Estimates with Fewest Remaining Issues
8.3 Overall Summary of Identified Issues
Issue 1. "Minor Systematic Error"
Issue 2. "Substantial Systematic Error"
Issue 3. "AV Crash Selection Bias"
Issue 4. "AV MVMT Mileage Selection Bias"
Issue 5. "Unrecorded CV Crash Selection Bias"
Issue 6. "CV MVMT Mileage Selection Bias"
Issue 7. "CV Crash Rate Differences at AV versus CV Sites," Site Selection Bias
Issue 8. "Non-CA AV Test Sites," Site Selection Bias
Issue 9. "AV Companies Heterogeneity ," Best Practice Violation
Issue 10. "Crash Severity Heterogeneity ," Best Practice Violation
Issue 11. "CV Sites Heterogeneity ," Best Practice Violation
References
CHAPTER 9 Overall Discussion
9.1 Final AV IRR Estimates
9.2 Final AV IRR Issues
9.3 New Findings
9.4 Limitations of the Reanalyses
9.5 Crash Rate Reduction: AVs versus Crash Avoidance Technologies
9.6 Did Lower AV Speeds Reduce the AV Severe Crash Rate?
9.7 Other Hypotheses to Explain the Upward Trend in AV Cumulative Crash Rate
Hypothesis 5. Increased ODD Range
Hypothesis 6. Increased AVs
Hypothesis 7. CVs Are the Problem, not AVs
References
CHAPTER 10 Overall Conclusions
References
CHAPTER 11 Specific Recommendations
References and Endnotes
Appendix A AV Hopes and Modeling Studies
A.1. AV Advocate Hopes Are Based on Reducing Human Error
A.2. AV Safety Hopes Are Based on an Implicit Syllogism
A.2.1 Error 1: The Syllogism's Major Premise Is Misstated
A.2.2 Error 2: The Syllogism's Conclusion Is Overgeneralized
A.2.3 Error 3: The Syllogism Contains an Oversight
A.3. Modeling Studies
Appendix B Abbreviations and Definitions
Appendix C Transition Crashes: Descriptions and Comments
C.1. Do Test Drivers Cause Crashes When They Transition to Manual Mode?
C.1.1 Waymo AV/Motorcycle Crash on October 19, 2018
C.1.2 GM Cruise AV/Motorcycle Crash on December 7, 2017
Appendix D Should Crash Reports by AV Companies Be Accepted at Face Value?
Appendix E Why Unreported and Unrecorded CV Crashes?
E.1. Unreported CV Crashes
E.2. Non-PR and Unrecorded CV Crashes
E.2.1 Errors in Quoting Unreported CV Crash Data
E.2.1.1 Systematic Errors in Abstract
E.2.1.2 Systematic Error in Percentage of Unrecorded Injury Crashes
E.2.1.3 Systematic Error in Percentage of Total Unrecorded Crashes
E.2.1.4 Wrong Number of Households
E.3. Review of Unrecorded CV Crashes
E.4. Unrecorded AV Crashes
E.4.1 Unrecorded AV Crash on July 1, 2015
E.4.2 Unrecorded AV Crash on August 20, 2015
E.4.3 Unrecorded AV Crash on February 4, 2016
E.5. Discussion of Appendix E
Appendix F Are AVs or Human Drivers at Fault in a Crash?
F.1. Review of Fault Claims
F.2. Fault Issues
F.3. Discussion of Fault Issues
F.4. Recommendations about Fault Issues
F.5. Legal Questions regarding AV Fault
Appendix G Crash Severity Definitions in Naturalistic Driving Studies
Appendix H Standardization Method
H.1. Rationale for AV Crash Site Standardization
H.1.1 Google AV Severe Crash Rate Reduction Is a True Effect
H.1.2 Google AV PDO Crash Rate Reductions Are Not True Effects
H.1.3 Benefits of Standardization
H.2. Standardization Calculation Method
H.2.1 AV Severity Level 1 (Severe) Crash Standardization Factor
H.2.2 AV Severity Level 2 (Major PDO) Crash Standardization Factor
H.2.3 AV Severity Level 3 (Minor PDO) Crash Standardization Factor
Appendix I Replication of Teoh and Kidd (2017) Phase 1 CV Crash Results (Table 12, Note f)
Appendix J Crash Reduction: AVs versus Crash Avoidance Technologies
Appendix K Site Selection Bias: Site Differences in CV Crash Rates
K.1. Mountain View versus U.S. CV Total Crash Rate
K.2. Mountain View versus San Francisco CV Total Crash Rate
K.3. San Francisco versus U.S. CV Total Crash Rate
Appendix L Summary of Issues
Appendix M Do ADS Neural Networks Have Negative Learning?
M.1. "Overdetermined" ADS?
M.2. "Underdetermined" ADS?
M.3. Neither Overdetermined nor Underdetermined ADS?
Appendix N PR Crash Selection Bias
N.1 PR Crashes
N.2 Police-Recorded Crashes
N.3 PR Crash Selection Bias Varies Inversely with Crash Severity Level
N.4 Effect on Google AV Standardized Crash Rates
N.5 Discussion
N.6 Future Research Recommendation
Appendix Bibliography
References
Epilogue
About the Author
Index

System requirements

Save as PDF Copy link into clipboard

Schweitzer Fachinformationen

Critical Analysis of Prototype Autonomous Vehicle Crash Rates

Description

More details

Other editions

Additional editions

Content

System requirements