MD-TLCF: Miner Distance Detection Based on Trajectory-Based Low-Confidence Filter

The confidence level of the detection result generally indicates the reliability of the detected object. However, in underground coal mines with low light intensity and uneven light distribution, the confidence level of detection results is low and exhibits significant fluctuation. Conventional 2-D miner detection methods are inadequate in delivering comprehensive 3-D miner positions to coal mine robots. This article proposes a real-time detection framework on RGBD images to generate a miner detection box and corresponding distance (miner state). Moreover, a trajectory-based low-confidence filter (TLCF) is proposed to refilter the miner state with a low confidence value based on the predicted state of the tracked miner trajectory. To evaluate the performance of distance detection methods in underground coal mines under various lighting conditions, a miner tracking dataset MINERTKRGBD is proposed. In the evaluation of this dataset, the proposed framework based on RGBD images outperforms the state-of-the-art vision methods. The miner distance detection method using TLCF (MD-TLCF) outperforms the method that solely relied on a confidence threshold filter (CTF). The Recall50 of the MD-TLCF method improved by approximately 4.2% compared to the latter, without any decrease in Precision50. Our code repository and dataset are publicly available at
https://github.com/HT-hlf/MD-TLCF.git
.