A lightweight SG-YOLOv5s network model was proposed to solve the problems of complex structure, large number of parameters and difficult deployment on the end side of the current detection model of hazardous objects under vehicle. Firstly, the backbone and the neck of YOLOv5s network were optimized and improved, which significantly reduced the number of network parameters and greatly reduced the weight volume of the model. Secondly, in the training stage, Mixup data was used to enhance the generalization ability of the model. Finally, SIoU was used to replace the frame regression loss function CIoU, so that the hazardous object prediction box was closer to the real box and the detection accuracy was improved. In view of the fact that there were few data sets of hazardous objects under vehicle, a large number of fragmented images of car undersides were captured using an intelligent small car, and the AutoStitch algorithm was adopted to splice the images, and finally the self-built data set of car bottom images was obtained. Experimental results show that, the identification accuracy rate of SG-YOLOv5s model was 97.63% in the self-built data sets of nine simulated vehicle hazards, which was 1.26% higher than that of the original YOLOv5s model. Additionally, the SG-YOLOv5s model reduced the number of parameters by 71.27% and decreased the model weight volume by 71.28%. These advancements provide the potential for embedded deployment of subsequent recognition models.