【摘要】 文中基于显著性特征的多视角动作图像识别方法，自动学习并提取出运动员动作的关键特征，有助于教练为运动员制定更科学、更个性化的训练计划。将人体骨架序列对齐到统一的时空坐标系中，计算距离图和角度图以捕捉骨架的空间特征，生成人体运动特征图；构建CNN+CA模型，将处理后的多视角动作视频帧生成感兴趣区域（ROI）拼接图，再将其输入到CNN中，提取多视角融合特征，并在CA模块中突出那些对于动作图像识别最为关键的区域；通过序列匹配算法将多视角动作识别问题转化为预测标签序列的匹配问题，为待识别动作图像分配动作类别标签，实现准确的多视角动作图像识别。实验结果表明：该方法不仅能够有效处理来自不同视角的动作图像，还能够准确识别出篮球运动员的多种动作。更多还原

【Abstract】 A multi view action image recognition method based on saliency features is studied. The key features of athlete actions are automatically learned and extracted, which can help coaches develop more scientific and personalized training plans for athletes. The human skeleton sequence is aligned to a unified spatiotemporal coordinate system, and both the distance map and angle map are calculated to capture the spatial features of the skeleton, so as to generate a human motion feature map. A CNN+CA model is built. A region of interest(ROI) splicing diagram from the processed multi view action video frames is generated and input into CNN, and then multi view fusion features are extracted. In addition, the most critical region for action image recognition in the CA module is highlighted. By sequence matching algorithm, the multi view action recognition is transformed into a matching of predicting label sequences, and action category labels are assigned to the action images to be recognized, which achieves accurate multi view action image recognition. The experimental results show that the method can not only effectively process action images in different perspectives, but also accurately recognize multiple actions of basketball players.更多还原