MineDet: A Real-Time Object Detection Framework Based Neural Architecture Search for Coal Mines

Coal is the primary source of energy worldwide. In longwall mining, autonomous coal mining machines can assist or replace human beings in dangerous mining tasks to achieve safe and efficient production. However, the abnormal hydraulic support guard plate state poses potential safety risks. Therefore, it is imperative to utilize target detection in video surveillance to monitor the status of the hydraulic support guard plate in real-time. Meanwhile, traditional video surveillance systems suffer from latency and cost problems because they must transmit video data to a server for processing. To address these challenges, we propose MineDet, a real-time object detection model tailored for the hydraulic support status in fully mechanized longwall mining. MineDet offers faster speeds and maintains comparable average precision by combining Neural Architecture Search (NAS) with the YOLOV8 model. Specifically, an efficient search space is engineered via the reparameterization branch structure, and a binarized path search algorithm is devised, significantly reducing search time and memory usage. It is worth noting that we achieved the MineDet model in just 30 h of search time on a single GeForce RTX 4090 GPU. Compared to YOLOV8n, the MineDet model demonstrates a reduction of 35% in model parameters, a 15% decrease in inference latency, and a 25% decrease in FLOPs while experiencing only a marginal decline of 0.006 in mean Average Precision (0.5-0.95).