Spatial Perception algorithm for open-pit mining areas based on lightweight stereo matching

Spatial perception is a core capability of autonomous driving technology in mining, enabling the system to navigate and avoid obstacles in complex mining environments. Existing LiDAR-based ranging technologies are costly, sparse in data, and have poor real-time performance. The key challenge in open-pit mine autonomous driving systems is to reduce equipment costs and improve response speed while maintaining perception accuracy. To address this, a lightweight stereo-matching framework is proposed for efficient and accurate real-time stereo matching. A lightweight convolutional module is used as the backbone network, with a multi-scale convolutional attention module that extracts stereo image feature maps across three different scales. A skip connection-based upsampling module is used to construct a 3D cost volume. Based on this, depthwise separable convolutions are employed to reduce the parameter scale of traditional convolutions. A funnel-shaped module, which includes both downsampling and upsampling, processes the 3D cost volume to capture and aggregate multi-scale contextual semantic features, decoding high-resolution geometric details. In the downsampling module, the spatial features from the matching process are aggregated, expanding the receptive field and reducing computation. In the upsampling module, high-resolution texture details in the feature maps are restored. Finally, disparity regression is used to estimate continuous disparity maps. To acquire training data, a large-scale training dataset with dense disparity labels was constructed using LiDAR’s sparse point cloud ground truth as a supervision signal, based on the AutoMine open-pit mining dataset. Through various benchmark experiments against state-of-the-art methods, the disparity estimation error was reduced by 45.6%, inference time was reduced to 17 ms, and computational cost was lowered to 42.5% over other methods, proving its superiority in speed, accuracy, and resource efficiency. This demonstrates the effectiveness of the proposed convolution block channel optimization method. The lightweight stereo matching network achieved the highest accuracy and lowest latency in the lightweight stereo matching method on the AutoMine dataset, showing progress in real-time stereo matching for complex mining environments. This advancement enables the deployment of the algorithm on embedded in-vehicle devices, promoting the application of stereo matching algorithms in spatial perception and navigation obstacle avoidance in open-pit mine autonomous driving.