Impact of Neutron-Induced SEU in FPGA CRAM on Image-Based Lane Tracking for Autonomous Driving: From Bit Upset to SEFI and Erroneous Behavior
Tomonari Tanaka, Wang Liao, Masanori Hashimoto, Yukio Mitsuyama
Abstract
Configuration random access memory (CRAM), which consists of the static random access memory susceptible to single event upset (SEU), configures all of the user logic in field-programmable gate array (FPGA). In this article, we evaluate the impact of SEU in CRAM on the image-based lane tracking for autonomous driving via neutron irradiation experiments. In the experiments, the cross section of bit upsets and its corresponding single event function interrupt (SEFI) in the logic function of image processing are measured. By using a virtual driving environment, we observe whether a bit upset finally induces a severe SEFI of the system failure. The system failure is observed by an abnormal behavior on the virtual autonomous car driving by lane tracking. Experimental results show the bit upset has a maximum probability of 23% to induce SEFI and finally 8% of the bit upsets lead to failures in lane tracking. All the SEFIs observed in irradiation experiments are reproduced in fault injections with the same bit address. The cross section of SEFI was estimated in fault injections with reasonable precision. Moreover, we evaluate the improvement of system reliability by the error correction against soft errors. Our evaluation result shows that the error correction within the period of two-frame processing time could reduce the 74% SEFI cross section, but it has little benefit for system failure cross section if an error happens during the period of the image processing.