V-BSDST: a Vehicle Blind Spot Detection and See-Through Method Based on V2V Perspective Sharing

Vision, as a critical medium for vehicle perception, serves as the most intuitive approach for acquiring environmental information during driving. However, existing vehicle vision systems are limited by fixed viewpoints and constrained perceptual ranges, rendering them incapable of addressing blind spot issues caused by external occlusions. To tackle these challenges, this paper proposes V-BSDST, the first vehicle blind spot detection and see-through method based on vehicle-to-vehicle perspective sharing. To overcome the limitations of single-vehicle fixed viewpoints, we introduces multi-viewpoint sharing to achieve complementary perception information. To precisely locate blind spots and occlusion areas, we integrates large-scale visual understanding models with state-of-the-art image segmentation models. For the seamless integration of images from different viewpoints, we presents an adaptive feathering fusion algorithm. Furthermore, to enable real-time operation of the algorithm on a hardware simulation platform, we designed a dual-thread parallelization method and introduced inter-frame constraints based on transformation matrix propagation. Extensive experiments conducted on a custom-built real-world dataset and physical scenarios demonstrate that V-BSDST outperforms existing image inpainting and stitching algorithms.