摘要
由于存在人体之间相互遮挡、受环境变化影响较大、时序特征提取能力较弱等问题,现有的动作识别方法在精准度方面仍有不足,因此,提出了一种基于改进Slow-Only网络的骨骼点动作识别方法。首先,将骨骼关键点数据进行预处理,分别在时间和空间维度减少冗余信息;其次,基于Slow-Only网络,重新设计了时间卷积模块,以更好地提取视频帧所包含的时序信息;最后,增加了改进的注意力机制模块,以降低遮挡问题带来的影响。在NTU RGB+D数据集上进行了实验,实验结果表明该方法能有效地提升检测精度,并且在实际场景中具有应用价值。
Due to the problems of mutual occlusion between human bodies,great influence by environmental changes,and weak ability of temporal feature extraction,the existing action recognition methods still have deficiencies in accuracy.Therefore,a skeleton point action recognition method based on improved Slow-Only network is proposed.Firstly,the skeletal key point data is preprocessed to reduce redundant information in time and space dimensions respectively.Secondly,based on the Slow-Only network,the time convolution module is redesigned to better extract the timing information contained in the video frame.Finally,an improved attention mechanism module is added to reduce the impact of occlusion problems.Experiments are carried out on the NTU RGB+D dataset.
出处
《工业控制计算机》
2023年第7期54-57,共4页
Industrial Control Computer
基金
国家重点研发计划资助(2020YFB1600702)。
