基于改进YOLOX的落石检测方法

Rockfall detection method based on improved YOLOX

摘要:

山坡地区是落石频发的区域，凭人力难以及时发现灾害的发生。为及时检测到落石的发生并做出应对措施，提出一种基于改进YOLOX的落石检测方法，自动检测并报告落石的发生情况；通过自制落石数据集训练YOLOX网络，优化空间金字塔池化结构，获取更多语义信息，并引入ECA-Net(Efficient Channel Attention Module，高效通道注意力模块)，提高特征的提取能力和特征间的信息传播，同时改进损失函数并使用数据增强，提高网络训练效果；实验结果表明，改进YOLOX算法的mAP@0.5为92.50%，每秒检测帧数为62.6，相较于YOLOX算法，mAP@0.5提高3.45%，每秒检测帧数上涨0.3；与原算法相比，在不损失性能的情况下，精度有较大的提升，同时满足图片与视频数据的实时检测要求。

Abstract:

Hillside areas are prone to falling rocks, so it is difficult to detect the occurrence of disasters in time by manpower. In order to timely detect the occurrence of falling rocks and take countermeasures, a method of falling rocks detection based on improved YOLOX is proposed to automatically detect and report the occurrence of falling rocks. The self-made rockfall data set is used to train YOLOX network, optimize the spatial pyramid pool structure, and obtain more semantic information. The attention mechanism of ECA-Net(Efficient Channel Attention Module) channel is introduced to improve the feature extraction ability and information transmission between features. Meanwhile, the loss function is improved and data enhancement is used to improve the network training effect. The experimental results show that mAP@0.5 of the improved YOLOX algorithm is 92.50%, and the number of frames detected per second is 62.6. Compared with the YOLOX algorithm, mAP@0.5 is 3.45% higher and the number of frames detected per second is 0.3 higher. Compared with the original algorithm, the accuracy is improved greatly without loss of performance, and the real-time detection requirements of image and video data are met.