基于MK-YOLOV4的矿区人员无标注视频检索方法

Unlabeled Video Retrieval Method of Mining Personnel Based on MK-YOLOV4

对矿区人员入矿、出矿及重要生产活动行为进行精准定位与准确识别是实现矿区智能安全生产的重要基础。针对复杂的矿区生产环境,提出一种基于MK-YOLOV4的矿区人员无标注视频检索方法,实现对矿区重要关口监控无标注视频的多人员目标检测和各人员身份重识别。首先,提出MK-YOLOV4算法,实现无标注视频多人员检测,在YOLOV4上构建多尺度预测,结合K-means++算法聚类生成符合样本特点的anchor box,增强卷积神经网络对小目标的表征学习。其次,提出基于外观不变性的通道注意力特征提取网络,实现矿区人员身份的精确重识别,并针对矿区人员统一工作服的难点,提出基于权重约束的难样本采样损失函数,结合Color jitter和随机擦除两种数据增强策略,提高身份识别网络的精确性和鲁棒性。最后,针对现有训练数据集类别少、场景样本单一等特点,构建了更符合矿区场景特点的Miner-Market矿区人员重识别数据集,并在标准数据集和该数据集上对所提方法进行验证,充分证明了所提方法具有较高的检索性能和识别精度。

Precise positioning and accurate identification of personnel entering, exiting, and conducting important production activities in the mining area are important foundations for achieving intelligent and safe production in the mining area. This study proposes an unlabeled video retrieval method for personnel in the mining area using MKYOLOV4 in the complex mining area production environment, which can realize multiperson target detection and reidentification of an individual’s identity on unlabeled video of important gateway monitoring in the mining area.First, this study proposes the MK-YOLOV4 algorithm to achieve multiperson detection of unlabeled videos by building multiscale predictions on YOLOV4, and the K-means++ algorithm is combined to generate an anchor box that meets the characteristics of the samples, which can improve the representation learning of the convolutional neural network for small targets. Second, we propose a channel attention feature extraction network based on appearance invariance to achieve accurate reidentification of personnel in mining areas. Aiming at solving the problem of uniform work clothes for personnel in mining areas, this study proposes a weight-constrained difficult sample sampling loss function with two data enhancement strategies, where Color jitter and random erasure are combined to improve the accuracy and robustness of the identification network. Finally, according to the characteristics of the existing training dataset with few categories and single scene samples, a Miner-Market mining personnel reidentification dataset is constructed with the characteristics of the mining scenes, and the proposed method is verified on the standard dataset and Miner-Market dataset.The verification confirmed that the proposed method has high retrieval performance and recognition accuracy.