Object Detection Based on Multi-View Images to BEV Feature

With the development of autonomous driving technology, environment awareness has become the key to ensure the safety of autonomous driving. Currently, bird's eye view (BEV) has become a popular research direction for object detection under BEV features because it can provide a panoramic view to help vehicles accurately perceive their surroundings, but the traditional methods cannot effectively utilize multi-view image data, and the low-cost image data still suffer from the problem of inaccuracy of view transformer. Images can offer rich texture and semantic message, but lack direct depth message, which affects the accuracy of BEV feature representation. Traditional view transformer methods rely on geometric projections, which have limited effect in complex scenes such as object occlusion and road undulation. For this reason, this paper innovatively proposes PerConBEV, a multi-view images BEV detection model based on deep learning, by improving the view transformer module, combining depth and semantic information, and using the probabilistic sampling module to aggregate to obtain BEV features. Finally, CenterPoint detection head is utilized for object detection under BEV features. Empirical analysis shows that the method in this paper presents significant performance advantages in both single-frame input and time-sequenced multi-frame conditions The research content of this paper provides a new perspective and methodological basis for BEV detection in multi-view images.