UTD-YOLO: a novel YOLO for uncovered truck detection

Due to the diverse truck loading types and varied shapes of goods carried by trucks on the highway, accurately judging whether a truck is covered with a tarpaulin presents challenges. Additionally, due to the large size of trucks, occlusion issues arise, which significantly degrade detection performance. To address these challenges, this paper presents a novel YOLO-based Uncovered Truck Detection method (UTD-YOLO). Firstly, we incorporate a Squeeze Aggregated Excitation Auxiliary Information (SAEAI) module into the auxiliary branch, which focus on crucial feature information related to uncovered trucks during the feature fusion process and assists the main branch in reducing semantic information loss. Secondly, we introduce a Large Separable Kernel Attention Spatial Pyramid Pooling (LSKASPP) module into the backbone network to enhance the model's capacity to learn diverse features of uncovered trucks while maintaining lower params. The proposed method is validated using real highway freight vehicle data, and the experiment results demonstrate that UTD-YOLO achieves outstanding performance with a precision of 82.37% for mAP@0.5 and 62.84% for mAP@[0.5:0.95]. Notably, mAP@[0.5:0.95] surpasses state-of-the-art object detection methods including YOLOv9, PPYOLOE, RT-DETR, YOLOv8, YOLOv6, and YOLOv5 by 2.17%, 2.68%, 2.8%, 2.83%, 9.63%, and 5.19% respectively. UTD-YOLO also demonstrates significant advantages in model size and inference speed, making it more suitable for smart highway roadside edge applications. This method contributes to improving the warning capabilities for highway safety risks, thereby advancing the field of Intelligent Highway Traffic Monitoring System.