Vehicle Object Detection Based on Improved RetinaNet
Luyang Zhang, Haitao Wang, Xinyao Wang, Shuai Chen, Huaibin Wang, Kai Zheng, Hailong wang
Abstract
Abstract Aiming at the low efficiency of vehicle object detection in real scenes, this paper proposes an improved RetinaNet. An octave convolution structure and a weight pyramid structure are introduced respectively to improve the detection performance of RetinaNet for vehicles. Specifically, we use octave convolution instead of the traditional convolution layer to improve the feature map’s representation of detailed information. In addition, in order to improve the quality of feature fusion, a weighted feature pyramid network (WFPN) structure is proposed to limit the propagation of gradients between different levels. The experimental results on the DETRAC data set show that the method has good detection results for vehicle targets of different scales in different scenarios, and can better meet the needs of practical applications.