Environment Influences on Uncertainty of Object Detection for Automated Driving Systems

Possible causes for malfunctioning behaviour in the Safety Of The Intended Functionality (SOTIF) are closely related to object detection algorithm for the functionality under development. In order to analyse uncertainty of object detection for Automated Driving Systems (ADS), we provide a new approach to systematically analyse the influence of environmental factors on object detection with Average Precision (AP). We generate test cases depending on real-world changes driving environment like dark, sunset, rain, snow, etc. This paper uses a Recurrent Rolling Convolution (RRC) model which is an outstanding Deep Neural Network (DNN), it can improve the accuracy of vehicle while realizing the real-time object detection. In particular, by using AP, this paper evaluated our model under different environment and found thousands of erroneous detection many of which may lead to potentially safety risk in the top-performing DNNs. The method proposed in this paper can help to improve system design, verification and validation work for the SOTIF.