In order to solve the problems of low detection accuracy, frequent target ID switching and limited terminal deployment of existing road target detection algorithms, a deep neural network-based road multi-target detection algorithm is proposed. In the detection stage, Ghost convolution is integrated into YOLOv5s backbone network to reduce the computational volume, SimAM (Simple Attention Module) attention mechanism is introduced to improve the perception of small targets in complex environments, Focal EIoU loss function is used to replace CIoU in the original network to improve the convergence speed, and Soft Non-Maximum Suppression ( Soft Non-Maximum Suppression (Soft-NMS) is used to replace the traditional Non-Maximum Suppression (NMS) to solve the problem of target occlusion and missed detection, and finally, the local vehicle-pedestrian detection dataset is self-constructed by LabelImg to train the network and improve the generalization ability of the model. The optimized model is deployed to the Jetson Nano development board and equipped with a real vehicle for validation, and the results show that the improved YOLOv5s improves the mAP50 by 2.7% and the processing speed by 8.3% based on a 47.4% reduction in the amount of parameters. The results of this study can provide a reference for multi-target tracking of intelligent driving vehicles.