Research of Highway Vehicle Inspection Based on Improved YOLOv5

Vehicle detection has become an important detection target for highways, but the traditional vehicle detection technology has poor real-time performance and large model parameters. The algorithm is based on YOLOv5, which introduces the improved network structure Ghostnet-C in the backbone layer to simplify the network structure and while increasing the detection speed of it. Subsequently, for further optimize the structure of the model, GSConv + Slim-neck structure is introduced in the neck layer. Finally, the CAS attention mechanism is used in the neck layer, which is developed in this paper, to change the focus of model predictions and get better results from the model. Compared with original YOLOv5, W-YOLO we propose in the paper reduces the amount of parameters by about 58.6%, the size of storage space by 58%, and the computation by 72.2%, while the accuracy can reach 75.6%. From the final results of experiments, we can discover that W-YOLO can significantly reduce the amount of parameters, model size and computation while guaranteeing accuracy, which can satisfy the requirements of highway vehicle detection more easily.