Ensuring SOTIF: Enhanced object detection techniques for autonomous driving

Abstract

Neural networks’ insufficient interpretability can lead to unguaranteed Safety of the Intended Functionality (SOTIF) issues when perceptual results are not always met in autonomous driving applications. To address the safety shortcomings in the current object detection process, this study proposes an object detection algorithm to enhance the accuracy of the perception system’s detection. We utilize the classical one-stage object detection algorithm YOLO v5 as the baseline in this study and evaluate our proposed model. A prediction extension box is added to the classical YOLO v5 model, which considers the coverage range and redundancy of real targets, guaranteeing the safety of image perception. The proposed object detection algorithm has been shown to increase the coverage range of detected targets, which significantly enhances perception safety in the autonomous driving process.