Behavior Decision-making Algorithm For Solving Safety Of The Intended Functionality Of Intelligent Vehicles

The complex and changeable external environment,system performance limitations,driver misuse,and other factors cause traffic accidents of intelligent vehicles.Under such non-fault conditions,the system function is not as good as expected / the unreasonable risk of hazard events caused by driver misuse is the problem to be solved by Safety of the Intended Functionality（SOTIF）.Due to the lack of interpretability and uncertainty of its network model,the end-to-end decision-making algorithm is commonly used in automatic driving based on deep learning,leading to the prominent problem of SOTIF.This thesis proposes an improved algorithm to enhance the interpretability of this kind of decision algorithm and estimate the uncertainty to improve the SOTIF evaluation.The research contents of this thesis are as follows:（1）Current intelligent vehicle end-to-end decision algorithms have the problem of insufficient interpretability.This thesis proposes a new multi-view attention module,which can predict the intelligent vehicle’s speed and steering wheel angle by establishing the relationship between the targets in different views.The experimental results on the drive360 dataset show that the visualization of the attention Heatmap of the multi-view attention module significantly improves the model’s interpretability and explains the decision logic of most scenes.At the same time,the decision accuracy of the algorithm is improved by 15.1% compared with the latest surround method.（2）Current intelligent vehicle end-to-end decision algorithms have the problem of uncertainty estimation.This thesis proposes a method of using a deep ensemble to measure the uncertainty of the end-to-end decision-making network.By integrating multiple multi-view attention decision-making networks,complete the uncertainty estimation of the end-to-end decision-making network,and predict the intelligent vehicle speed and steering wheel angle simultaneously.The experimental results show that the network can accurately output the uncertainty in various complex traffic road scenes and improve its decision-making accuracy by 3.5% based on research content（1）.（3）In order to verify the system performance of the proposed algorithm in actual scenarios,this thesis builds a hardware and software platform for actual vehicle acquisition to collect campus data sets.An offline experiment is employed on a lightweight model for verifying the decision accuracy,interpretability,and uncertainty performance in the campus scenario.The experimental results prove the effectiveness of the decision-making algorithm with uncertainty estimation in unique campus scenarios（e.g.,speed bumps,dense crowds.）and greatly enhanced the interpretability of some complex scenarios.