A fused network based on PReNet and YOLOv4 for traffic object detection in rainy environment

In order to improve the detection accuracy of vehicle target in severe rainy day under traffic environment, a deep learning network DTOD-PReYOLOv4 (derain and traffic object detection-PRcNct and YOLOv4) was proposed based on the fusion of PReNet and YOLOv4, which integrated the improved image restoration subnet D-PRcNct and the improved target detection subnet TOD-YOLOv4. D-PRcNct could extract rain streak features more effectively, since it introduced the multi-scale expansion convolution fusion module (MSECFM) and the attentional mechanism residual module (AMRM) with SEBlock into PReNet. TOD-YOLOv4 improved not only the detection accuracy of small traffic target, but also the detection efficiency, since it replaced the backbone module CSPDarknct53 of YOLOv4 with the lightweight CSPDarknct26 of YOLOv4, added CRB into PANet of YOLOv4 neck, and utilized k-means++ instead of the original network clustering algorithm. DTOD-PRcYOL()v4 was verified based on the constructed vehicle target database VOD-RTE in rainy day traffic scenario. Research results show that compared with the current scries of YOLO networks, the proposed DTOD-PRcYOLOv4 can better extract the features with lower resolutions by superimposing RB over RcsBlockbodyl in the shallow layer. It can effectively reduce the convolutional layer redundancy and improve the memory utilization, since ResBlock_body3, ResBlock_body4 and RcsBlock_body5 in deep layer can be properly cropped to RcsBlock_body3 × 2, ResBlock_body4 × 2 and RcsBlock_body5 × 2, respectively. It also can alleviate the degradation of small target detection effect caused by the deepening of network layers by adding jump connection to Concat+ConvX 5 in PANet to form CRB. In the process of multi-scale detection, k-means-++ algorithm is adopted to allocate smaller prior boxes that are more suitable for the larger feature images, but larger prior boxes that are more suitable for smaller feature images, which further improves the accuracy of target detection. The harmonic mean value of precision and recall rate, average precision and detection speed of DTOD-PReYOLOv4 respectively increase by 5.02%, 6. 70% and 15. 63 frames per second compared with MYOLOv4, by 3. 51%, 4.31% and 2.17 frames per second compared with TOD-YOLOv4, by 46.07%, 48.05% and 18.97 frames per second compared with YOLOv3, and by 31.06%, 29.74% and 16.26 frames per second compared with YOLOv4.4 tabs, 12 figs, 44 refs. © 2022 Chang'an University. All rights reserved.