Lightweight target detection and multi target tracking for UAV inspection in open pit mines

Abstract

Aiming at the problems of low efficiency in traditional manual inspection of open-pit mines, difficulty in identifying faulty equipment and non-cooperative targets, high safety risks, and insufficient detection accuracy for small targets, while supplementing the limitations of active positioning technologies such as UWB indoor-outdoor positioning and vehicle-mounted strapdown inertial navigation in scenarios like signal blind areas and non-cooperative target monitoring, this paper proposes a lightweight object detection and multi-target tracking algorithm, and constructs an intelligent UAV inspection system. In the design of the detection model, deformable convolution DCNv2 is introduced into the backbone network, and the progressive feature pyramid network AFPN is adopted in the neck part to enhance the multi-scale feature extraction capability. A lightweight detection head (LSDECD-Head) is designed, and combined with the Focaler-GIoU loss function, the detection accuracy of small targets and occluded targets is improved. The LAMP pruning algorithm is used to compress the model, and under a 30% pruning rate, the model still maintains a performance of mAP50 at 0.868 and inference time of 196 ms, which is suitable for the computing resource constraints of UAVs. In terms of multi-target tracking, the ByteTrack algorithm is improved. A space-appearance similarity matrix (ASM) that integrates the target’s spatial position, operation status, and appearance features is introduced, and combined with an acceleration correction function to optimize trajectory prediction. This improvement increases the multi-target tracking accuracy (MOTA) by 2.6% and reduces the number of ID switches by 21. In addition, a multi-level inspection system is constructed, which integrates functions of data collection, real-time detection, and multi-UAV collaborative scheduling. It realizes data transmission and remote monitoring relying on 5G and ad-hoc network technologies. The core innovation of this paper lies in constructing the C2f-DCN+AFPN lightweight feature extraction architecture, tailored to capture complex target features in mining areas. Designing the LSDECD-Head detection head and Focaler-GIoU loss function to enhance difficult sample detection. Proposing a Hierarchical Adaptive LAMP Pruning Strategy to Balance Accuracy and Lightweighting. Enhanced ByteTrack algorithm incorporates ASM matrix and acceleration correction to improve dynamic tracking stability: Establishing an air-ground collaborative inspection system to achieve technological implementation. The aforementioned innovations are not merely a simple combination of existing technologies, but rather a deeply integrated optimization addressing the pain points specific to open-pit mining scenarios. Experimental results show that this scheme significantly improves the accuracy and stability of equipment detection and tracking in open-pit mine scenarios, and provides a feasible technical solution for intelligent and unmanned inspection of mines.