Abstract: To address the high deployment cost and limited disaster resilience of downhole wireless positioning systems, an active downhole target positioning method is proposed based on an improved ConvNeXt network and panoramic vision technology. The method consists of two stages: Offline training and online positioning. In the offline training stage, the downhole environment is divided into regions, and panoramic image data are collected to construct a multi-scale training dataset through data augmentation. Based on the improved ConvNeXt network, an area positioning model and a fine positioning model are trained separately, where the area model provides coarse-grained position estimation and the fine model achieves precise localization. During the online positioning stage, a panoramic camera mounted on a mobile device continuously captures environmental images. After image stitching and enhancement preprocessing, the area positioning model first determines the target’s regional range, and the corresponding fine positioning model is then invoked for real-time localization. Experimental results demonstrate that the combination of 0.5 m area positioning and 0.1 m fine positioning achieves an accuracy of 94%. With the integration of the CoordAttention module, the accuracy further increases to 95%, representing a 14.67% improvement over conventional single-step methods. Compared with MobileNetV3 (91%) and ResNet (92.33%), the improved ConvNeXt network exhibits superior performance. Furthermore, the proposed method maintains strong robustness against typical downhole disturbances such as illumination variation, motion blur, personnel movement, and dust interference, achieving 95.33% accuracy on the 616 roadway dataset collected from the Yangchangwan Coal Mine of the National Energy Group Ningxia Coal Industry Company Limited. The study provides a low-cost, robust, and easily deployable positioning solution for intelligent downhole mobile platforms such as autonomous vehicles and robots.