The keypoint algorithm, as a fundamental algorithm in machine vision, plays a crucial role in enhancing the visual perception capabilities of new mining equipment. The keypoint algorithm can be applied across various mining tasks. The unique characteristics of the mine environment, such as lighting variations, dust interference, lack of environmental texture, and repetitive texture structures, present significant challenges for keypoint algorithms. To effectively evaluate the applicability of keypoints in underground mine environments, this paper constructed two types of datasets. The first dataset was the mine coal wall image test dataset, containing 20 sets of challenging coal wall or tunnel wall image sequences, while the second was the tunnel inspection image dataset, recording 589 image frames from a wheeled robot during an inspection process. In comparative experiments, we evaluated various keypoint algorithms, including SIFT, ORB, SURF, AKAZE, L2-Net, HardNet, GeoDesc, SuperPoint, R2D2, and DISK. The experimental results show that deep learning-based keypoint algorithms exhibit superior overall performance, with R2D2 demonstrating significant advantages over other algorithms. Additionally, we evaluated the efficiency of deep learning-based keypoint algorithms on low-power edge computing platforms, further validating their feasibility in industrial applications.

To achieve the intelligent recognition of coal gangue in fully mechanized caving face, a coal caving acoustic signal collection device is designed, which can sense the movement of the tail beam and automatically trigger data collection.Field data is collected at the fully mechanized caving face 3106 of Gucheng Coal Mine, Shandong Energy, and it is manually labeled to construct a sample library of acoustic signal classification for top coal caving.Then, six machine learning classification methods are applied in the time domain, frequency domain and time-frequency domain, and the classification effect of them are evaluated by different frame lengths and different feature vector dimensions.The results show that: the classification effect based on time-frequency domain features is the most stable, and its accuracy rate is the highest by different frame lengths.The classification accuracy rate of random forest, K-nearest neighbor, decision tree and multi-layer perceptron model is above 80 %.Among them, the classifier performance based on wavelet packet decomposition and random forests are the best, and the classification accuracy is 93.06 %.There is a correlation between the time-frequency domain feature vectors and higher dimensions.Through dimensionality reduction, a small number of comprehensive features can be extracted and the amount of system calculations can be reduced.The principal component analysis is used to reduce the time-frequency domain feature vector to 20.Thus, the classification accuracy rate is further improved to 94.51 %.