A New Method of Non-Deep Network for Fast Vehicle Detection

To achieve object detection on low-computing devices such as embedded and mobile devices, we propose a new method of the non-deep network for fast vehicle detection. In our method, we use color transformation to address the problem of insufficient training data. To achieve effective object detection of different sizes, we introduce a non-deep network with a parallel double-stream design. The upper stream adopts the downsampling block which contains a 3*3 kernel size convolution layer to extract the feature of small objects. The lower stream uses the downsampling block which contains a 5*5 kernel size convolution layer to realize the detection of large targets. Finally, these extracted features of different receptive fields are fused in the fusion block. We use the one-level output feature from backbone for detection to improve model efficiency. The experimental results show that our network runs real-time on Jetson TX2, and achieves 30.46% mAP on COCO and 77.2% mAP on UA-DETRAC. Our detector is more accurate than YOLO-Fastest and faster than YOLOv4-tiny.