Underground personnel detection and tracking using improved YOLOv7 and DeepSORT

The coal industry is undergoing a transformation in the concept of intelligent mining with "safety, efficiency, intelligence, and green" as its core. Computer vision, as an emerging technology with high efficiency, intelligence, and low cost, has become an important highlight in the current construction of intelligent mines. The underground monitoring video is susceptible to interference factors such as artificial light source, dust and spray, which leads to the poor real-time performance, high missed & false detection rates and poor tracking accuracy of existing underground personnel monitoring methods using computer vision. Based on this, an improved YOLOv7 and DeepSORT underground personnel detection and tracking algorithm is proposed. First, in order to be able to extract more critical underground personnel image features and improve the model's adaptability in the complex scene of coal mine underground, the SimAM attention mechanism is incorporated into the Neck module of YOLOv7, and the improved YOLOv7 model is used to detect underground personnel targets. After that, in order to be able to further improve the tracking accuracy of personnel targets while reducing the number of model parameters and network complexity, the ShuffleNetV2 lightweight module is introduced into the feature extraction network of DeepSORT, and the improved DeepSORT model is used to encode and track downhole personnel targets. Finally, the experimental validation of the proposed algorithm is carried out on the established downhole personnel video image dataset and the public dataset, and the results show that: the average detection accuracy of the improved YOLOv7 model is improved by 3.9% compared with that of the YOLOv7 model; the multi-target tracking accuracy of the improved DeepSORT algorithm reaches 74.9%, the tracking accuracy reaches 82.3%, and the speed reaches 24 FPS. Compared with YOLOv7-DeepSORT algorithm, the number of network parameters of the improved algorithm is reduced by 36%, significantly improving the real-time performance of underground multi personnel target detection and tracking, and is expected to be deployed in the underground intelligent edge computing monitoring platform.