SYG-Net: A New High-Precision Vehicle Detection Network

Abstract

To improve the accuracy of vehicle detection, a vehicle detection neural network SYG-Net with YOLOv3 network as the main body was proposed and combined with generalized Intersection over Union (GIoU) and spatial pyramid pooling (SPP) module. The backbone network of SYG-Net network is the basic network structure of YOLOv3. However, a layer of SPP was added before the main network structure of feature extraction, namely, darknet and YOLO layers. In this manner, the features before the input of YOLO layer can obtain spatial features. GIoU was used as the regression loss of BBox at the end of the network layer and tested on UA-DETRAC data set. Results showed that the map and recall values of SYG-Net network increased substantially. Meanwhile, loss and average GIoU converged quickly and had good effect. SYG-Net was 0.75% and 0.75% more accurate than YOLOv3 and 0.7 YOLOv3-SPP, respectively. Results showed that SYG-Net was effectively detects vehicle. This paper looks forward to the combination of SYG-Net and other modules.