DTS-YOLOv8n: A Lightweight Vehicle And Pedestrian Detection Model

In target-intensive detection environments, the detection of vehicles and pedestrians faces the problem of misdetection and omission of occluded objects and small target objects, as well as the challenge that the model is too complex to be lightweight. In this paper, the YOLOV8n model is studied and improved, and a lightweight vehicle pedestrian detection network DTS-YOLOv8n is proposed. Firstly, by introducing the C2f_DWR_DRB module, the reparameterised large convolutional kernel layer improves the performance without increasing the computational cost, which enhances the detection speed and accuracy of the model, and secondly, the Triplet Attention mechanism is added to improve the performance of the model under the lightweight architecture, and in order to further reduce the complexity of the model and to maintain the detection accuracy, the study also combines the GSConv and VoV-GSCSP modules and introduces the Slim-Neck method, thus achieving more efficient and accurate vehicle and pedestrian keypoint detection. The experimental results on KITTI dataset show that the accuracy rate is increased to 89.3%, the recall rate to 74.3%, [email protected] increased to 84.3%, corresponding to a 0.6% improvement, and [email protected] also increased to 56.7%, corresponding to a 0.3% improvement. Meanwhile, the number of parameters and the calculation amount of the model are reduced by 13.3% and 13.6% respectively. While reducing computational costs, the detection accuracy was significantly improved, and all performance metrics outperformed the original YOLOv8n model.