Research on the Application of YOLOv8 Model Based on ODConv and SAHI Optimization in Dense Small Target Crowd Detection

In view of the wide application of dense small target population detection technology in many fields such as traffic management, emergency response, and commercial applications, it is necessary to improve the accuracy and timeliness of relevant models. In order to solve this problem, this paper proposes a new lightweight and efficient dense small target population detection algorithm based on YOLOv8: ODConv2-YOLO-SAHI. The ODConv dynamic convolution module is embedded into the backbone network of YOLOv8, which enhances the adaptability of the model to the diversity of input data and reduces the computational cost of the model. The ODConv multi-dimensional attention mechanism was introduced to improve the model's ability to extract individual features of small targets by learning complementary attention in multiple dimensions along the kernel space. In the process of model inference, the slicing aided hyper inference mechanism (SAHI) was introduced to improve the small target detection capability of the model. Experimental results show that ODConv2-YOLO with only ODConv dynamic convolution module and ODConv multi-dimensional attention mechanism improves the detection accuracy by 2.1 % compared with YOLOv8n, while the parameter quantity and computational amount are only 2.8M and 7.3GFLOPs; After adding SAHI mechanism, the performance of the ODConv2-YOLO-SAHI model is improved by about 25% compared with YOLOV8n in the average number of detected small target individuals, which better meets the needs of high detection accuracy and low computing power for dense small target crowds in field scenarios.