Abstract: Coal block detection methods mainly include traditional image detection methods and deep learning target detection methods. The traditional image detection method has low detection precision and poor real-time performance, and can not accurately determine the coal pile. Although the deep learning target detection method can achieve real-time detection, it does not identify the number, retention, and blockage of coal blocks. And there are many identification model parameters. To solve the above problems, a coal block abnormal behavior identification method based on improved YOLOv5s + DeepSORT is proposed. Firstly, video images of coal blocks on a belt conveyor in a fully mechanized coal mining face are collected by the camera and inspection robot, and data sets are made. Secondly, the MobileNetV3_YOLOv5s_AF-FPN model is used for detecting the coal image target. The original YOLOv5s backbone feature extraction network is replaced by MobileNetV3 to reduce the number of parameters and improve the reasoning speed. The original feature pyramid network in YOLOv5s is improved to AF-FPN to improve the detection performance of the YOLOv5s network for multi-scale coal targets. DeepSORT is used for multi-target tracking of coal blocks. The coal block image detected by the improved YOLOv5s is taken as the input of DeepSORT for multi-target tracking. DeepSORT is used to estimate the state of coal blocks, perform data association and matching, and update the tracker parameters to determine the tracking results. The continuously tracked coals are ID-coded, and the number of coals in the current frame is counted. Finally, the continuously tracked target is taken out from the target tracker, and a distance threshold is set. Whether the target is detained or not is determined. The quantity threshold is set to determine whether it is blocked. The identification of abnormal behavior of coal block retention and blocking state is finally realized. The reliability of the coal abnormal behavior identification method based on the improved YOLOv5s + DeepSORT is experimentally verified by using the self-built dkm_data2021 data set. The results show that compared with the YOLOv5s model, the average detection precision of the improved YOLOv5s model is improved by 1.45%, the parameter quantity is reduced by 35.3%, the reasoning is accelerated by 12.7%, the average missed detection rate is reduced by 11.08%, and the average false detection rate is reduced by 11.54%. The detection precision of coal block abnormal behavior identification method based on the improved YOLOv5s+DeepSORT is 80.1%, which can accurately identify the status of coal block retention and blockage. The result verifies the reliability of the method.