Enhancing Safety Perception in Autonomous Driving Systems Through 3D Object Detection and Neural Network Regression
Yuhang Wang, Yue Zhang, Shutong Liu, Chuangqi Li · 2024
With the rapid development of autonomous driving technology, safety perception has become a critical component in ensuring the reliability and safety of these systems. However, challenges persist, such as traffic accidents caused by the Ideal L9, which misidentified billboard images as real vehicles, leading to erroneous emergency stops. This paper proposes a novel method to address such safety perception issues by enhancing 3D object detection accuracy. Utilizing the mmdetection framework and the SparseBEV algorithm, we identify and address deficiencies in distinguishing planar images from stereo objects. Our approach combines deep neural network regression learning with geometric constraints, introducing a hybrid discrete-continuous loss function to improve the 3D direction regression of objects. Experimental results demonstrate significant improvements in detection accuracy and safety awareness, providing a robust solution to mitigate misidentification risks. However, further optimization is needed to ensure comprehensive safety enhancement in autonomous driving systems.