In this paper, we propose a practical and reliable car accident sound recognition model using deep learning techniques. In this study, 400 car accident sound files were collected, segmented, and classified into 1312 sound files for training the model and 327 sound files for testing the model. The sound files were transformed into spectrograms using a short-time Fourier transform. YOLOv7 was utilized to train the model to detect the sounds of vehicle skidding and collisions. During the model training, image augmentation parameters need to be turned off so that the trained overall model can achieve an average accuracy of 0.875 for vehicle skidding and collision sounds in car accidents during testing. The threshold for average precision was set to 0.8, and the misdetection rate for common vehicle horn sounds was kept below 22.5%. The verification results of this car accident sound detection model demonstrate its practical application capability.