ZZ-YOLOv11: A Lightweight Vehicle Detection Model Based on Improved YOLOv11

Aiming at the problems of insufficient vehicle detection accuracy, high misdetection and omission rate, and heavy model computational burden caused by complex lighting conditions, target occlusion, and other factors in urban traffic scenarios, this paper proposes an improved lightweight detection network, ZZ-YOLO. Firstly, the current mainstream target detection algorithms lack components to improve the network's focus on the edges of the objects, which can indirectly lead to unclear classification and localization. For this reason, in this paper, we self-develop a module of GlobalEdgeInformationTransfer (GEIT), which can help us to transfer the edge information extracted from the shallow features to the whole network and fuse it with the features of different scales. Secondly, to reduce the number of parameters in the detection head and to fuse the extracted features better, a self-developed Lightweight Detail Convolutional Detection Head (LDCD) detection head is introduced. After that, the most effective layer-adaptive magnitude-based pruning (LAMP) method is used to build away the redundant parameters to make the detection network more lightweight. Finally, in order to ensure that the detection accuracy of the pruned model will not be too low, a model distillation method was used, in which YOLOv11x + LDCD was used as the teacher model and the pruned model was distilled as the student model. Experimental data on the optimized KITTI and BDD100K datasets show that the detection accuracy of the ZZ-YOLO algorithm is 70.9%, the mAP (mean Average Precision) @0.5 is 58%, the model-parameter quantity is 14.1GFLOPs compared to the original algorithm, the detection accuracy is increased by 5.7%, and the average precision is increased by 2.3%. The amount of model parameters is reduced by 34%, and the real-vehicle verification session effectively reduces the misdetection and omission of vehicles.