为了解决在背景相似的篮球视频中提取特征级运动信息不充分和捕获长时序依赖关系困难等问题,从局部和全局的角度出发,提出一种混合运动激励和时序增强网络(mixed motion excitation and temporal enhancement network,MTE-Net),该网络由在时间建模上互补的混合运动激励(mixed motion excitation,MME)模块和时序增强(temporal enhancement,TE)模块构成.混合运动激励模块通过计算短距离视频帧之间混合的特征级差分来充分表征局部运动信息,并显性地对运动敏感通道进行激励.时序增强模块对长距离视频帧使用自注意力机制来构建时序关联函数并捕获时序之间的全局依赖关系,增强视频中的重要帧序列.在不额外引入光流和过多参数的情况下,在SpaceJam篮球动作数据集上的实验结果表明,与其他主流的动作识别算法相比,所提模型对篮球运动员动作识别的准确率更高.
Action recognition algorithm for basketball player based on mixed motion excitation and temporal enhancement
To solve the problems of insufficient extraction of feature-level motion information and difficulty in capturing long temporal sequence features in basketball videos with similar backgrounds,this paper proposes mixed motion excitation and temporal enhancement network(MTE-Net)from a local and global perspective.The network is composed of mixed motion excitation(MME)module and temporal enhancement(TE)module,which are complementary in time modeling.Firstly,the MME module fully characterizes local motion information by calculating mixed difference between short distance video frames,and explicitly stimulates motion-sensitive channels.Secondly,the TE module uses self-attention mechanism to con-struct temporal sequence correlation function for long distance video frames and capture global dependence between temporal sequences to enhance important frame sequences in videos.Finally,without adding additional optical flow and excessive pa-rameters,experimental results on SpaceJam show that compared with other mainstream action recognition algorithms,the model proposed in this paper has a higher accuracy for basketball player action recognition.
deep learningaction recognitionmotion featuretemporal enhancement
国家科技期刊平台
NSTL
万方数据
