Diagnostic Mechanism and Robustness of Safety Relevant Automotive Deep Convolutional Networks

Abstract

Convolution Neural Networks today provide the best results for many image detection and image recognition problems. The computational complexity and the amount of parameters learned has increased, yet there is little to no research on the topic of functional safety for systems incorporating CNNs. The analysis on false detections due to random hardware faults concentrates on human made adversarial examples obtained by adding unrealistic noise sources over carefully selected images. Redundant execution of these networks is prohibitive in application domains where power and price constraints dominate, pushing for alternate approaches. In this paper we investigate functional safety aspects for a road labeling application, a common task in the advance driver assistance systems. We introduce computationally light safety checks that reduce the error space significantly, train a CNN on the Cityscape dataset that reaches 93% mean IU (intersection over union) and use Monte Carlo simulations to assess the impact of single event upset random hardware faults. The results show that the networks based on convolution and ReLU (rectified linear unit) have some intrinsic robustness and that together with additional constraints strong function safety claims can be made. We compare also the diagnostic coverage between floating point and fixed point implementation of CNNs and summarize key safety features needed to achieve a high diagnostic coverage.