The YOLOv8m road object detection algorithm based on a lightweight design

Abstract To address the issues of low detection accuracy, slow detection speed, and high computational resource consumption in current road object detection algorithms, a lightweight YOLOv8m road object detection algorithm is proposed. Firstly, a lightweight convolution, GSConv, is introduced, and a C2fGSConv module is specifically designed to substitute the Conv and C2f modules within the Neck layer. This substitution effectively decreases the complexity of the algorithm during the feature fusion stage, thereby enhancing the network’s inference speed. Secondly, a lightweight convolution partial convolution (PConv) is introduced to redesign the detection head of YOLOv8, further reducing the model’s parameter size. Lastly, the efficient attention mechanism is integrated into the Backbone layer to enhance the algorithm’s detection accuracy for small road objects. Experimental results on the BDD100K road object detection dataset show that the improved algorithm achieves a 0.5% increase in detection accuracy, a 31.39% reduction in model computation, a 27.4% improvement in detection speed, and outperforms other comparison algorithms in various evaluation metrics, meeting the requirements of road object detection tasks.