Abstract: The underground spatial structure of metal mines is complex, and the images captured from different angles can yield varying positions for the same target, leading to discrepancies in target positioning and negatively impacting the accuracy of personnel tracking. This study proposed a method for personnel tracking and positioning in underground metal mines through image feature correction. The underground images were preprocessed using local adaptive Gamma correction. Light distribution features were extracted from the images using gray histograms, scatter plots, and discrete cosine transforms. The optimal Gamma value was then determined using the K-nearest neighbor algorithm to achieve uniform lighting in the images. The study introduced a Gaussian mixture model for foreground segmentation, effectively separating the target from the background by dynamically updating the background model, which enhanced target recognition accuracy. During the tracking and localization phase, homography matrix decomposition and vanishing point calculation were utilized to map multi-view images onto a reference plane, resulting in a multi-layer fused image. This process corrected the geometric distortions between images captured from different angles, ensuring that the positional information of the same target was consistent across various views, thus achieving preliminary localization of multiple targets. Furthermore, graph cut theory was incorporated to construct an energy function and optimize trajectory association using a temporal sliding window, ultimately enabling robust localization and continuous tracking of multiple targets in the underground environment. Validation experiments demonstrate that the application of this method effectively enhances the information entropy and clarity of the images, with tracking and localization errors maintained within 10 pixels. This allows for precise tracking and positioning of personnel in underground metal mines, thereby ensuring their safety.