VRU Detection Based on Inverse Perspective Mapping and Deep Learning in Lightweight Embedded System

This paper presents a novel approach for real-time detection and warning of vulnerable road users (VRU) in lightweight embedded systems, enhancing the efficiency of Bird's Eye View (BEV) image stitching. By exploiting the geometrical characteristics of image mapping, a rapid BEV transformation method is proposed, aiming at effectively detecting VRU in front of the vehicle. Central to this approach is the innovative combination of image distortion and inverse perspective mapping into a unified map through a specially designed map fusion algorithm. This fusion not only streamlines the BEV stitching process but also aligns the size of the EBV image with the input requirements of the detection model, thereby reducing preprocessing time and computational load. Empirical results demonstrate that the proposed method significantly reduces time consumption by approximately eight times compared to traditional method, marking a notable advancement in the field of real-time VRU detection for intelligent driving systems.