【摘要】 针对复杂场景下手势分割准确性低,手势细粒度特征描述不充分和手势识别实时性差的问题,提出融合批量再标准化和YOLOv3的手势识别算法。首先,在复杂背景及不同光照条件下采集20种手势,运用数据增广策略进行样本扩充并建立标准手势库;然后通过K均值维度聚类获取训练集手势锚点框,负责对不同尺度手势进行检测;最后利用迁移学习和微调方法训练得到手势识别模型。为解决YOLOv3网络在手势训练阶段和预测阶段进行规范化处理时数据间可能存在较大偏差问题,采用批量再标准化方法提高手势识别准确性。手势识别过程具有快速、准确、非接触的优势,实验表明在正常实验环境下,手势平均识别率为97.6%,对于复杂背景下干扰较大的手势平均识别率达到89.2%以上,单次手势识别速度为0.04 s。更多还原

【Abstract】 Aiming at the problem of low accuracy of gesture segmentation,inadequate description of fine-grained features and poor real-time gesture recognition in complex scenes,a hand gesture recognition algorithm combined with batch renormalization and YOLOv3 is proposed. Firstly,20 gestures are collected under complex backgrounds and different lighting conditions,the data augmentation strategies are used for sample expansion,and a standard gesture library is established. Then,the anchor is obtained through K-means dimension clustering to detect the different scale hand types;Finally,the gesture recognition model is obtained by using the transform learning and fine-tuning. In order to solve the problem that there may be large deviations between the data when YOLOv3 network are standardized in the gesture training phase and the prediction phase,the batch renormalization method is adopted to improve the accuracy of hand gesture recognition. Experiments show that the average accuracy in normal experimental environment is 97. 6%. the average recognition rate of gestures is over 89. 2% in the complex environment,and the recognition speed is 0. 04 second.更多还原