摘要:
智能视频监控已被广泛应用于煤矿井下安全管控环节,而人员危险行为检测是保障煤矿安全生产和建设智慧矿山的重要内容。受煤矿井下复杂工况、设备遮挡、光照不均及粉尘等因素影响，现有目标检测算法在人员危险行为检测时存在样本提取困难、特征提取不准确等问题，导致检测精度低且鲁棒性不足。为此，本文提出一种基于YOLOv8的煤矿井下人员危险行为检测方法PMR-YOLO，以提高检测精度和鲁棒性。该方法主要包括基于部分卷积（Partial Convolution，PConv）的特征提取增强，基于混合局部通道注意力机制（Mixed Local Channel Attention，MLCA）的特征捕捉增强和基于感受野注意力卷积（Receptive-Field Attention Convolution, RAFCAonv）的识别能力增强三个部分。首先，采用PConv替换标准卷积识别图像遮挡内容，引入掩码进行自适应识别和填充，提高模型的抗遮挡能力。然后，针对数据集中特征过多而无重点偏好的问题，在特征提取环节引入MLCA，融合空间和通道信息进一步模型表征能力和检测性能。此外，采用RAFCAonv替换原有算法的检测头模块，将空间注意力和感受野特征相结合，增强模型对不同尺度和复杂度图像的适应性。实验结果表明：相比于YOLOv8模型，PMR-YOLO模型在准确率、精确度和召回率分别提高了6.3%、13.7%和10.1%。该方法显著提高了煤矿井下人员危险行为检测精度，为复杂工况环境中的人员危险行为检测提供了新的思路。
Abstract:
Intelligent video surveillance has been widely used in underground coal mine safety control, and personnel risky behavior detection is an important part of coal mine safety production and the construction of intelligent mines. Affected by the complex working conditions, equipment shielding, uneven illumination and dust in underground coal mines, the existing target detection algorithms in the detection of dangerous behaviors of personnel have problems such as sample extraction difficulties and feature extraction inaccuracies, resulting in low detection accuracy and lack of robustness. For this reason, this paper proposes a YOLOv8-based method PMR-YOLO for detecting dangerous behaviors of people underground in coal mines, in order to improve the detection accuracy and robustness. The method mainly includes feature extraction enhancement based on Partial Convolution (PConv), feature capture enhancement based on Mixed Local Channel Attention (MLCA) and feature capture enhancement based on Receptive-Field Attention Convolution (RAFCA). Attention Convolution (RAFCAonv) based recognition enhancement in three parts. First, PConv is used to replace the standard convolution to recognize the occluded content of the image, and a mask is introduced for adaptive recognition and filling to improve the occlusion resistance of the model. Then, to address the problem of too many features in the dataset without focus preference, MLCA is introduced in the feature extraction link to fuse spatial and channel information to further model characterization ability and detection performance. In addition, RAFCAonv is used to replace the detection head module of the original algorithm, combining spatial attention and sensory field features to enhance the model's adaptability to images of different scales and complexity. The experimental results show that compared with the YOLOv8 model, the PMR-YOLO model improves the accuracy, precision and recall by 6.3%, 13.7% and 10.1%, respectively. The method significantly improves the precision of detecting hazardous behaviors of people underground in coal mines, and provides a new idea for detecting hazardous behaviors of people in complex working environments.
下载:
