基于关键帧节点自适应分区与关联的行为识别算法

基于视频的人体行为识别任务中由于大部分画面并不包含重要的判别信息，这对识别应用的准确性造成严重干扰。关键姿态帧既能表达视频又能降低计算量，且骨骼数据相比于图像包含更多维度的信息。因此，提出一种基于关键帧骨骼节点自适应分区与关联的行为识别算法。首先构建自适应池化深度网络以评估帧的重要性获取关键姿态帧序列；其次通过节点自学习模型建立非自然连接状态下的节点间关联；最后将改进的时空信息应用于STGCN并使用softmax分类识别。在开源的大规模数据集NTU-RGB+D和Kinetics 上与几种典型技术进行比对，验证了所提方法在减少冗余数据量的同时能保留关键动作信息，且动作识别准确率平均提高了0.63%~11.81%。

Action recognition based on adaptive partition and association of key-frame nodes

In the task of human behavior recognition, most of the video frames do not include important discrimination information, which seriously affects the accuracy of application. Key pose frames can effectively express the video and reduce the amount of computation. Furthermore, bone data contains richer information than RGB image. Therefore, this paper proposed an action recognition approach based on adaptive partition and association of key-frame nodes. Firstly, it constructed an adaptive pooled deep network to evaluate frames importance and obtain key pose sequence. Then, it established association between nodes in unnatural connection state by self-learning model. Finally, it applied the improved spatio-temporal information on STGCN and used softmax for classification. This paper evaluated the effectiveness of the proposed approach by comparing with several typical technologies on the open-source and large-scale datasets of NTU-RGB+D and Kinetics. Experimental results show that it can reduce the amount of redundant data and retain key action information, and obtain higher average accuracy by 0.63% ~ 11.81% than the compared methods.