Research on personnel re-recognition method in coal mine underground based on improved metric learning

In the traditional personnel re-recognition method in coal mine underground based on metric learning, because metric learning ignores the absolute distance between positive and negative samples, the gradient of the loss function disappears or disperses. This results in low recognition precision of underground personnel position information. In order to solve this problem, a personnel re-recognition method in coal mine underground based on improved metric learning is proposed. Firstly, a feature extraction method for underground personnel based on manual design features is adopted to manually process and extract features such as color space and texture space, enriching the feature dimensions. Secondly, Euclidean distance is used to calculate the similarity of high-dimensional features of personnel. Finally, an improved triple loss function is proposed. Adding adaptive weights to the traditional triple loss function increases the weight of effective samples. It solves the problem of gradient disappearance or dispersion caused by ignoring the absolute distance between positive and negative samples. The traditional recognition method is compared with the personnel re-recognition method in coal mine underground based on improved metric learning for cumulative matching feature curve verification and recognition rate verification. The results show the following points. ① The personnel re-recognition method in coal mine underground based on improved metric learning has a sample matching probability of 100% when the number of similar samples is around 50. ② The personnel re-recognition method in coal mine underground based on improved metric learning reduces the inference time of two different calibration size images by 44 ms and 68 ms, respectively, compared to traditional re-recognition methods. ③ The personnel re-recognition method in coal mine underground based on improved metric learning performs better after discarding the images of personnel heads and feet. It has a sample matching probability of 100% when the number of similar samples is around 42.