面向无人门店的顾客行为检测研究

Research on Customer Behavior Detection in Unmanned Stores

摘要:

针对无人门店俯视监控条件下人体目标尺度压缩、遮挡频繁及交互行为短时弱动态特性导致的漏检与误检问题，对复杂场景下顾客行为检测方法进行了研究，在YOLOv8主干网络中高层特征提取阶段引入受约束的GhostConv结构，在特征金字塔高语义特征层嵌入方向感知注意力机制，并结合MPDIoU边界框回归损失优化定位过程；以检测阶段输出的空间对齐人体特征序列为输入，构建基于膨胀因果卷积的轻量级时序卷积网络，设计受控时序感受野结构进行短时行为建模。UCF101、HMDB-51及自建无人门店数据集实验结果表明，该方法在复杂俯视与遮挡环境下提升了行为检测精度与稳定性，满足无人门店顾客行为检测的应用需求。

Abstract:

To address missed and false detections caused by scale compression, frequent occlusion, and short-duration weak-dynamic interaction behaviors under top-view surveillance conditions in unmanned retail stores, a customer behavior detection method for complex environments was studied. A constrained Ghost Convolution (GhostConv) structure was introduced into the mid-to-high layers of the You Only Look Once version 8 (YOLOv8) backbone. A direction-aware attention module was embedded into the high-semantic feature layers of the feature pyramid network. The Minimum Point Distance Intersection over Union (MPDIoU) loss was adopted to optimize bounding box regression. Spatially aligned human feature sequences generated by the detection stage were used as input. A lightweight Temporal Convolutional Network based on dilated causal convolutions was constructed, and a controlled temporal receptive field was designed for short-term behavior modeling. Experimental results on UCF101, HMDB-51, and a self-built unmanned retail dataset show that the proposed method improves behavior detection accuracy and stability under complex top-view and occlusion conditions, meeting the application requirements of customer behavior detection in unmanned retail stores.