基于CCL-YOLO的洗手动作检测算法

Hand Washing Action Detection Algorithm Based on CCL-YOLO

摘要:

针对现有YOLOv7模型对洗手动作检测精度低，受环境噪声影响大，相似动作易混淆等问题，提出了一种基于改进YOLOv7的CCL-YOLO目标检测算法。该算法通过引入自注意力机制EALN-CotAttention，以获取更多的上下文信息；使用CARAFE优化YOLOv7中最近邻插值方法的上采样，促进了高效的内容感知与特征重组，提高了对动作检测的精度；通过SPPFCSPC替换YOLOv7中的SPPCSPC，以参数量较少的卷积核得到了相同的感受野；采用了LADH-Dectect检测头替换YOLO网络中原始的耦合检测头，提高了模型性能，相比于普通解耦检测头，减少了网络参数和GFLOPs，进而提高了推理速度。在自制数据集上进行实验，结果表明所提算法平均精度提升到81.2%，相比于YOLOv7算法检测精度提高7.2%，精确率提高2.9%，召回率提高11%，更好满足洗手动作检测的实际需求。

Abstract:

To address the issues of low detection accuracy for hand-washing actions, significant environmental noise interference, and confusion between similar actions in the existing YOLOv7 model, an improved CCL-YOLO object detection algorithm based on YOLOv7 is proposed. This algorithm introduces the self-attention mechanism EALN-CotAttention to obtain more contextual information; it utilizes CARAFE to optimize the upsampling of the nearest neighbor interpolation method in YOLOv7, promoting efficient content perception and feature recombination, thereby improving the accuracy of action detection; it replaces the SPPCSPC in YOLOv7 with SPPFCSPC, achieving the same receptive field with fewer parameters in the convolutional kernel; it adopts the LADH-Dectect detection head instead of the original coupled detection head in the YOLO network, enhancing model performance. Compared to ordinary decoupled detection heads, this reduces network parameters and GFLOPs, thereby increasing inference speed. Experiments on a custom dataset show that the proposed algorithm improves average precision to 81.2%, which represents a 7.2% increase in detection accuracy, a 2.9% increase in precision, and an 11% improvement in recall rate compared to the YOLOv7 algorithm, better meeting the practical needs of hand-washing action detection.