Deep learning-based binocular vision for blast hole recognition and localisation

A deep learning-based binocular vision method for blast hole recognition was developed, specifically for the loading phase of open-pit blast operations. An enhanced YOLOv10 model was proposed, featuring integrated ELA and LSKA features through weighted fusion with an ‘alpha’ parameter for weight control. A C2f_LESKA module was constructed to improve feature extraction. The model’s performance was boosted through multi-level feature pyramids and bidirectional information transfer. EIOU was employed to address the slow convergence and imprecise localisation issues of the original algorithm. Experimental validation demonstrated the method’s efficacy in precise blast hole recognition and 3D positioning.