Negative obstacle detection on open pit roads based on multi-feature fusion

With the gradual implementation of intelligent mine concept, intelligence and unmanned operation are gradually implemented in mining area. Unmanned driving of open pit mine trucks is increasingly becoming the main focus of intelligent mine construction. In order to solve the safety problems of the overturn of unmanned vehicles and heavy-duty trucks due to irregular negative obstacles appearing in some parts of road surface such as potholes and collapses in open pit mines, and to improve the safe driving coefficient in mines, a multi-feature fusion method of detecting negative obstacles in open pit mine roads is proposed. The method uses the BiFPN feature fusion module to improve the weight proportion of small-scale negative obstacle detection, introduces the spatial and channel dual attention mechanism to improve the feature extraction and feature fusion ability of negative obstacle edges, so as to improve the detection accuracy of small-scale negative obstacles on the road. Also, the SIoU Loss is adopted as the loss function of the model bounding box, the Anchor by using the K-means++ method is used to improve the convergence speed and boundary frame localization effect of the obstacle detection model, the hyperparameters are optimized based on genetic algorithm to make the model more suitable for the mine scene, and finally the fast and accurate recognition of negative obstacles on the mine road is realized. The experiments show that the detection model can quickly and accurately identify the negative road obstacle targets in the complex background of the open pit mine, and the detection accuracy, recall rate, and mAP of the negative road obstacle targets reach 96.9%, 89.9%, and 95.3%, respectively, and the size of the model is only 12.7 MB. Compared with other mainstream detection networks, the network model is more suitable for the safety needs of unstructured road driving in open pit mining areas under complex environment, and the robustness of the detection model is good, which can be adapted to a variety of situations in open pit mining areas, providing a feasible method for the detection of negative obstacles on unstructured roads in open pit mining areas where the actual environment is complex and variable, and providing some safety warnings for the safety of unmanned trucks in open pit mines.