A real-time vehicle detection algorithm based on instance segmentation
Zhiyong Shan, Haiyun Tian · 2021
Vehicle detection has achieved great progress in intelligent transport system, which has the advantages of low risk and high accuracy. In this work, we adopt one-stage algorithm YOLACT to train our model on dataset MS COCO 2017. We propose (Precise) P-CIoU loss to better predict bounding box regression in vehicle detection. In terms of the final candidate bounding box selection, this paper proposes skip-calculation NMS mechanism, in which IoU threshold γ and normalized central point distance method incorporated. This mechanism can effectively solve similar occlusion problem. The experimental results show that our approach achieves performance improvement with gains about 2%, without sacrificing inference efficiency.