Abstract: With the continuous development of artificial intelligence technology and the gradual advancement of the smart mining concept, traditional open-pit mining operations are being replaced by automated methods. Unmanned mining trucks are a crucial component of smart mining. Their widespread adoption has effectively addressed safety issues caused by irregular terrain, uneven road surfaces, and adverse weather conditions, such as truck overturning and skidding, thereby significantly reducing the occurrence of accidents and fatalities.Accurate target category detection is a prerequisite for obstacle avoidance decisions, and model lightweighting facilitates deployment in resource-limited conditions. Therefore, to address the issues of high parameter count, large model size, and low detection accuracy for small and occluded targets in open-pit mining scenarios, we propose the Lightweight Unmanned Mining Truck Detection Algorithm LWHP (Lightweight High-Precision). This algorithm introduces the Efficient Weighted Bidirectional Feature Pyramid Network (R-BiFPN), which reconstructs the neck network through cross-layer connections and bidirectional propagation, reducing redundant computation paths. It enhances multi-scale feature fusion capability via weighted feature fusion, significantly reducing parameter count while improving small target detection capability. Additionally, it designs a detection decoupling head with a multi-head attention mechanism to improve the issue of network complexity caused by convolutional layer redundancy, processes spatial dimensions to focus on capturing target features, reduces interference from irrelevant backgrounds, and enhances the accuracy of occluded target recognition.urthermore, it constructs a lightweight neural network with dual convolution (CDC), enhancing inter-channel information flow, improving model feature expression capability, and reducing model complexity. The introduction of the Focal-EIOU loss function calculates the width and height differences of target bounding boxes and uses Focal Loss to address the imbalance of difficult and easy samples, achieving faster convergence and superior localization capability. Experiments show that the improved algorithm reduces parameters by 50.2%, computational load by 46.3%, and model size by 47.6% compared to the original algorithm, with a model size of only 3.3 MB. And the FPS reaches 92.9, meeting the real-time requirements.Accuracy is increased by 1.6%, recall by 3.1%, and average precision reaches 79.6%, a 2% improvement over the original model, ensuring lightweight deployment while enhancing detection accuracy.