Towards Safe Autonomous Driving
Doctoral thesis, 2021
In the first part of the thesis, methods for evaluating the safety benefit are investigated. In particular predictive effectiveness evaluation based on resimulation of accident data, using models to estimate new outcomes in case the safety system had been available. To illustrate the methodology, four examples of gradual increase in model complexity are presented. First, an Autonomous Emergency Braking (AEB) system using a sensor model, decision algorithm, vehicle dynamics model and regression based injury model. This is extended in a Forward Collision Warning (FCW) system which additionally requires a driver model to simulate driver reactions. The third example shows how an active, AEB, and passive, airbag, system can be combined. Finally the fourth example combines several systems to emulate a highly automated vehicle. Apart from predicting the real world performance, this analysis also identifies current safety gaps by studying the residual of the accident set.
Safety benefit estimation using accident data gives an evaluation on the current accident distributions, however, the systems may introduce new accidents if not operated as intended. In the second part of the thesis, safety verification processes with the intent of preventing unsafe operation, are presented. This is particularly challenging for machine learning based components, such as neural networks. In this case, traditional analytical verification approaches are difficult to apply due to the non-linearity and high dimensional parameter spaces. Similarly, statistical safety arguments often require unfeasible amounts of annotated validation data. Instead, monitor functions are investigated as a complement to increase safety during operation. The method presented estimates the similarity of the driving environment, compared to the training data, where decisions inferred from novel data can be considered less reliable. Although not providing a complete safety assurance, the methodology show promising initial results for increasing safety. In addition, it could potentially be used to collect novel data and reduce redundancy in training data.
effectiveness
Autonomous driving
neural networks
verification
predictive evaluation
monitoring
safety benefit
Author
Arian Ranjbar
Chalmers, Electrical Engineering, Systems and control
Overview of main accident scenarios in car-tocyclist accidents for use in AEB-system test protocol
Other conference contribution
Integrated bicyclist protection systems - potential of head injury reduction combining passive and active protection systems
24th International Technical Conference on the Enhanced Safety of Vehicles,;(2015)
Paper in proceeding
Predicted road traffic fatalities in Germany: The potential and limitations of vehicle safety technologies from passive safety to highly automated driving
Conference proceedings International Research Council on the Biomechanics of Injury, IRCOBI,;Vol. 2018-September(2018)p. 17-52
Paper in proceeding
Driving scene retrieval by example from large-scale data
CVPR Workshops 2019,;(2019)
Paper in proceeding
Scene Novelty Prediction from Unsupervised Discriminative Feature Learning
2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC),;(2020)
Paper in proceeding
Safety Monitoring of Neural Networks Using Unsupervised Feature Learning and Novelty Estimation
IEEE Transactions on Intelligent Vehicles,;Vol. In Press(2022)
Journal article
Arian Ranjbar, Nils Lubbe, Erik Rosen, Jonas Fredriksson. Car-to-pedestrian forward collision warning revisited: A safety benefit estimation
This thesis aim to look into the safety potential of autonomous driving, presenting methods of how the effectiveness and safety benefit can be predicted before the vehicles have entered the market. However, these methods only account for scenarios and accidents occurring in traffic today. To prevent the system from causing new types of accidents, verification methods are implemented, with the aim of assuring safe operation. This is particularly challenging for components using machine learning, due to the complexity and the vast amount of data required to cover all potential scenarios. The second part of the thesis aim to contribute towards the safety by using monitoring systems, comparing the driving environment to what the algorithms have previously seen in the training. Predictions made in novel environments are then considered to be less reliable. This in turn can be transmitted to the decision making of the vehicle, in order to perform conservative driving actions when provided with unreliable information, hopefully preventing potentially dangerous situations.
Subject Categories
Robotics
ISBN
978-91-7905-604-9
Doktorsavhandlingar vid Chalmers tekniska högskola. Ny serie: 5070
Publisher
Chalmers