In underground mines, high scene similarity and camera orientation variations pose challenges to autonomous visual positioning in tunnel environments. This study therefore proposes a hierarchical candidate localization method for underground visual positioning based on GoogLeNet and panoramic images. Specifically, a panoramic camera was used to capture 360° tunnel images with no blind spots. Environmental features were enhanced through preprocessing techniques such as image stitching and adaptive histogram equalization. We then constructed a hierarchical candidate localization framework combining partition recognition and location segmentation based on an improved GoogLeNet network. Multi-scale feature extraction was used in partition recognition to achieve efficient classification. A fine-grained model was then employed within the region to achieve precise location determination. Results showed that, by combining a partition interval of 0.7 m with a fine subdivision interval of 0.1 m, the system achieved a combined positioning accuracy of 84.33%, an increase of 34.00 percentage points over direct positioning. It could maintain its robustness even on extreme test sets, including complete darkness, flipping, and image acquisition during camera movement. Compared to AlexNet and VGG, GoogLeNet maintains high accuracy while requiring only 6.8M parameters. This significantly reduces computational overhead, thus making it suitable for embedded device-side deployment. This approach provides references for developing efficient, reliable, and deployable solutions for localizing personnel and equipment in underground coal mines.