基于轻量化YOLOv5n算法的交通目标检测研究OA

This paper proposes a lightweight and pruning strategy based on the improved YOLOv5n algorithm to address the trade-off between detection speed and model accuracy in traffic target detection.By analyzing and optimizing the architecture of the YOLOv5n model,an effective lightweighting strategy is developed to increase the computational speed while maintaining the model accuracy.First,the YOLOv5n network is lightweighted by introducing the GhostNet in the Backbone.The Conv modules in the Neck are replaced with GSConv to achieve performance comparable to standard convolution while reducing computational cost.Meanwhile,the bounding box regression loss in the Head is optimized using EIOU Loss.After training the improved model,pruning is applied to further reduce model size.The pruned model is evaluated on the dataset,improving the detection speed by 15 fps compared with the original model.It is further compared with the mainstream improvement methods.Finally,the improved model is validated on a mobile experimental platform.Results demonstrate with lightweight design and appropriate pruning,the proposed method markedly improves traffic target detection with an average fps of 130.34,up by 11.3%compared with the original model,while maintaining a detection accuracy of 83%with mAP@0.5.