基于改进YOLOv8的露天矿区行车障碍物检测

Traveling Obstacle Detection in Open-Pit Mine Area Based on Improved YOLOv8

露天矿区场景复杂,行车障碍物检测受扬尘和颗粒物等粉尘噪声干扰严重,难以准确识别障碍物,尤其是光线较差的夜间,不利于做出正确决策,从而影响无人作业的安全性和整体效率。针对以上问题,提出了一种基于YOLOv8n模型的露天矿区行车障碍物检测算法YOLOv8n-Enhanced。该算法主要从3个方面进行了改进,具体包括:首先,针对受粉尘噪声干扰严重和夜间光线不足的问题,提出了C2fCA模块结构,提高了模型特征提取能力;其次,使用轻量级卷积技术GSConv和VoV-GSCSP模块,减轻模型复杂性,实现检测器更高的计算成本效益;最后,使用WIOU损失函数,提高了模型泛化能力。试验结果表明:改进算法在保持实时性的前提下,可将YOLOv8n的平均精度(mean Average Precision,mAP)分别提高1.8%和2.6%,实现白天与夜间场景下不同尺度的障碍物识别。

The open-pit mining area is a complex scene,and the traveling obstacle detection is seriously interfered by dust noise such as dust and particles,which makes it difficult to accurately identify obstacles,especially at night when the light is poor,which is not conducive to correct decision-making,thus affecting the safety and overall efficiency of unmanned operation.In view of the above problems,a YOLOv8n-based YOLOv8n-Enhanced algorithm for detecting traveling obstacles in open-pit mining areas was proposed.The algorithm is mainly improved in three aspects:Firstly,for the problems of serious interference by dust noise and poor light at night,a C2fCA module structure was proposed instead of the original C2f module,which utilizes the shared weights and context-aware weights to enhance the dependency relationship between different locations of the image,mitigate the noise interference,and improve the feature extraction ability of the model.Secondly,to trade-off the accuracy and real-time performance of the open-pit obstacle detection model,the Neck end of the model was reconstructed,and the lightweight convolutional techniques GSConv and VoVGSCSP modules were used to reduce the complexity of the computation and network structure,and realize a higher computational cost-effectiveness of the detector.Finally,for the situation that there is a large gap between the quality of data in the open-pit mining area,especially at night when there is insufficient light,and lowquality data will affect the ability of the model to learn features in training,the loss function was optimized to solve the problem of the bounding box regression equilibrium between the samples of different qualities,to improve the ability of the model to generalize and accelerate the convergence.The experimental results show that the improved algorithm in this paper reduces the computational GFLOPs of the model by about 8.5%and the number of parametric params by about 3%while maintaining the real-time performance,and improves the mean Average Precision(mAP)of the YOLOv8n by 1.8%and 2.6%in daytime and nighttime scenarios,respectively,and realizes obstacle recognition at different scales in daytime and nighttime scenes.