To address the pose variations of target coal gangue caused by belt slippage, deviation, and speed fluctuations during coal gangue sorting operations, which make traditional belt-speed–based tracking methods incapable of accurately perceiving target poses in real time and may result in missed, failed, or empty grasps, this paper proposes a dynamic coal gangue tracking method based on an adaptive multi-feature ECO-FCC algorithm. An adaptive feature fusion strategy is employed to integrate CNN, fHOG, and CN features extracted from coal gangue tracking images, thereby improving target localization accuracy. An adaptive scale estimation strategy is introduced to dynamically adjust the tracking bounding box and obtain optimal target scale estimation. In addition, an occlusion-aware model updating mechanism is designed to enhance the robustness of the proposed method under complex operating conditions. Based on a self-developed dual-arm gantry-type coal gangue sorting robot platform, comparative tracking experiments are conducted under different belt speeds, different target scales, and different camera–robot installation configurations, in which the proposed method is compared with KCF, CN, fDSST, SiamFC, C-COT, ECO, and ECO-HC. Experimental results demonstrate that the proposed method achieves an accuracy of 97.4%, a success rate of 90.9%, and an average frame rate of 52.4 FPS, exhibiting high tracking precision, strong robustness, and good real-time performance, and effectively meeting the requirements for accurate dynamic tracking of coal gangue.