Road Target Detection Algorithm for Severe Rain and Snow Weather Based on YOLOv8s Improved Algorithm

Aiming at the current bad rainy and snowy weather road target recognition field detection of high computational complexity, low detection accuracy, easy to produce leakage, false detection and other problems, based on the YOLOv8s model to propose an improved road target detection algorithm in bad weather scenarios. Firstly, the Backbone layer of YOLOv8s structure is improved by presenting the efficient multi-scale attention module (EMA) to the feature fusion stage, which improves the target information retention ability and reduces the model computation, and enhances the model's ability to recognize occluded targets; Secondly, the Backbone convolutional layer is replaced with the SPD-Conv module with the aim of improving the problem of feature information loss of small targets during the convolution process; Thirdly, we add a parameter-free simple attention module (SimAM) after each C2f module in the Neck layer, which enhances the models capability for feature fusion, and in the meantime, we substitute the weighted bi-directional feature pyramid network (BiFPN) for the path aggregation network (PANet), achieving significant model benefits at the expense of minimal computational cost, and utilizing its ability to inhibit interference and enhance the attention of tiny targets; Finally, the Focaler-ShapeIoU loss function is employed to substitute the CIoU loss function of the initial model, thererby addressing the limitations of the bounding box regression method. Additionally, it can enhance the model convergence speed and bolster its target detection performance. Experimental validation conducted on the proprietary dataset demonstrates that the structurally optimized YOLOv8s algorithm yields a precision of 92.32 %, a recall rate of 90.67%, and an average precision of 93.14%, which is 3.05%, 2.16%, and 4.38% surpasses the performance of the initial YOLOv8s model, respectively, and the study results are better than other mainstream detection algorithms, for instance YOLOv5s and SSD, which demonstrates that newly version model has high recognition accuracy and detection speed, which significantly improves the road target detection accuracy in bad rainy and snowy weather.