Abstract: The automatic classification of mine car load based on convolution neural network in coal mine auxiliary shaft is realized in practical application. The misdetection and false alarm are caused by simple trigger conditions. Non-mine car objects passing through the detection area can cause the misoperation of driver controlled switch. In order to solve this problem, an intelligent identification method for mine car load in coal mine auxiliary shaft based on target detection model is proposed. An industrial camera is installed at the wellhead of a coal mine auxiliary shaft to collect the images of mine car load and the images are manually labeled so as to construct a mine car identification data set. And the identification accuracy and the real-time performance of three target detection models, namely Faster R-CNN, YOLOv4 and SSD are evaluated. According to the evaluation results, it is concluded that the YOLOv4 model is more suitable for the identification task of mine car load. In order to reduce the model size and improve the identification speed, the YOLOv4 model is improved. The lightweight network MobileNet is used to replace the original backbone characteristic extraction network CSPDarknet53. So the MobileNetv3-YOLOv4 model is constructed. The test results show that the mean average precision(mAP) of the MobileNetv3-YOLOv4 model is 95.03%, and the identification speed is 44 frames/s, which is 0.77% and 27 frames/s higher than that of the YOLOv4 model respectively. In order to facilitate field application and deployment and improve the performance of the mine car load identification model on the embedded platform, a model acceleration method based on inter-layer fusion and model quantization is proposed. The MobileNetv3-YOLOv4 model before and after the acceleration is transplanted to Jetson TX2 for field test of mine car load identification. The results show that the identification speed is increased from 18.3 frames/s before the acceleration of the MobileNetv3-YOLOv4 model to 35.42 frames/s, and the mAP is 94.68%, which meets the real-time and precise detection requirements in the field. And the detection task is only started when the mine car passes the detection area, which avoids the misoperation of driver controlled switch caused by other objects.