In-vehicle vision-based automatic identification of bulldozer operation cycles with temporal action detection

Abstract

Automated monitoring of bulldozer operation cycles is crucial for efficient productivity assessment and precise construction management. Harsh earthwork environments and complex, variable operation processes present challenges for identifying these cycles. To address this issue, we developed a multiscale temporal feature fusion and dual attention mechanism-based temporal action detection model (FDA-AFSD) for the automatic identification of bulldozer operation cycles from in to vehicle vision. This model enhances long-term sequence modeling, key temporal information learning, and precise action boundary identification through its multiscale temporal feature fusion structure, dual attention mechanism module, and scalable granularity perception (SGP) layer. In tests for earth levelling and mine edge dumping operations, the average detection accuracy (mAP) for bulldozer operation cycles reached 90.9%. Furthermore, under various adverse weather conditions and diverse construction processes, the model maintained stable and excellent detection capabilities, demonstrating its feasibility and practical application value.