Abstract: Mining engineering is rapidly developing towards automation and intelligence, and the construction of intelligent mines has become a future trend. The multi-objective environmental perception of unmanned mine cars in open-pit mine areas is a key step in unmanned transportation. For the safety risks caused by multiple obstacles such as rolling pits, puddles, vehicles and personnel in complex unstructured roads, the existing end-to-end algorithms face the challenges of small target information loss, insufficient multi-scale feature fusion, unbalanced samples, and difficulty in balancing model complexity and accuracy in the dynamic and complex environment of open-pit mines. To this end, an end-to-end multi-target detection model You Only Look Once-Mine Multi-target Detection (YOLO-MMD) for unmanned mine trucks in open-pit mines is proposed. For the problem of missing observation pixel information caused by unstructured terrain in open-pit mines, Space-to-Depth Convolution (SPD-Conv) is introduced to transform image spatial information into depth information, which effectively preserves the fine-grained perception ability of small targets in unstructured scenes and improves computational efficiency. In order to improve the effective use of context information, Efficient Multi-Scale Attention (EMA) is embedded in the detection layer to realize pixel-level cross-channel interaction and spatial information aggregation, which enhances the ability of multi-scale feature fusion without significantly increasing the computational burden. In addition, considering the sample imbalance problem of different obstacle target objects in open-pit mines, the In-Focaler-IoU loss function is designed to improve the efficiency and convergence speed of bounding box regression with auxiliary bounding box while paying attention to rare target samples. The study show that YOLO-MMD can detect multi-target objects under the conditions of occlusion and blurring, and achieve the best balance between multi-target detection accuracy and complexity. It can achieve 0.939 mAP, 4.56 MB model size and 5.8 G floating-point operations per second, which can provide effective and feasible environmental perception for the safe driving of unmanned mine cars.