School of Resource Engineering,Xi’an University of Architecture and Technology
Xi’an Key Laboratory of Intelligent IndustryPerception Computing and Decision Making,Xi’an University of Architecture and Technology
Xi’an U-Mine Intelligent ResearchInstitute Co.Ltd.

The emerging of autonomous driving for mining trucks in recent years has made obstacle detection critical.Unstructured roads in open pit areas are complex and variable, and small targets or multi⁃scale driving obstacles suchas stones and potholes often appear, which seriously endanger driving safety. Therefore, this paper proposes a methodfor detecting road obstacles in open⁃pit mines based on bidirectional feature fusion. The open pit barrier image datasetis further segmented and expanded through the field acquisition and usage of data augmentation methods, and aRepVGG+ backbone network structure more suitable for obstacle detection is proposed in the feature extraction stage. In the feature fusion stage, a bidirectional feature fusion pyramid model based on SimAM space with channel attentionand cross⁃stage connectivity is proposed. The performance of small target obstacle detection is improved by expan⁃ding the feature map and feature perception field for predicting small target obstacles, and the performance of multi⁃scale detection is improved by a bidirectional feature fusion mechanism. The convolutional layer and anchor size of the net⁃work classification and prediction module are also further adjusted to improve obstacle detection performance, reducefeature redundancy, and speed up model inference. In terms of the loss function of the model, for the problems of unbal⁃anced samples in training and imprecise obstacle bounding box localization, the Focal Loss with fused label smoothing regu⁃larization is used as the classification loss function, and the GIoU Loss is used as the bounding box loss function to furtheroptimize the open pit obstacle model. Experiments show that the method in this paper can effectively identify unstructuredroad obstacles in open pit mines in complex backgrounds. In practical applications, the detection accuracy reaches91.76% mAP and the detection speed reaches 56.76 fps, which has a better small target and multi⁃scale target detectionperformance compared with some mainstream detection methods, and can meet the requirements for the safe detection ofobstacles in the unmanned mine trucks traveling in open pit mines.