Search for a command to run...
Abstract Lipschitz continuity is a critical prerequisite for ensuring the object detection accuracy of unmanned autonomous driving systems in open-pit mining areas. However, significant order-of-magnitude disparity in Lipschitz constants between the restoration network and the detection network, which induces cascaded instability, thus amplifying detection errors and a decline in detection accuracy. In this paper, a Lipschitz Fusion Regularized CMT-3D (LFR-CMT-3D) framework is proposed, which upgrades the core paradigm from pure data restoration to the collaborative regularization of the input and parameter spaces. This framework embeds a Dust Adaptive Feature Calibration Fusion (DA-FCF) strategy, which estimates dust concentration based on the dark channel prior and dynamically adjusts fusion weights; a multimodal collaborative recovery network is constructed to achieve dust removal and geometric structure completion; we implement a dual-space Lipschitz regularization to constrain the sensitivity of the detection network to perturbations and mitigate gradient explosions. The proposed framework effectively addresses the problem of multimodal cascaded instability in the dusty environment of mining areas, mitigates cross-modal adaptation barriers, and can be seamlessly integrated into existing CMT-3D detectors. Experimental results on a self-constructed dataset under real open-pit mining conditions show that, compared with the original CMT-3D, the framework achieves a 22.26% improvement in mean Average Precision (mAP) under sparse, local and dense occlusion conditions, which significantly enhances detection stability and accuracy and provides reliable perceptual support for unmanned operations in mining areas.