Football Action Recognition using Hierarchical LSTM

被引:31
作者
Tsunoda, Takamasa [1 ]
Komori, Yasuhiro [1 ]
Matsugu, Masakazu [1 ]
Harada, Tatsuya [2 ]
机构
[1] Canon Inc, Tokyo, Japan
[2] Univ Tokyo, Tokyo, Japan
来源
2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION WORKSHOPS (CVPRW) | 2017年
关键词
D O I
10.1109/CVPRW.2017.25
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
We present a hierarchical recurrent network for understanding team sports activity in image and location sequences. In the hierarchical model, we integrate proposed multiple person-centered features over a temporal sequence based on LSTM's outputs. To achieve this scheme, we introduce the Keeping state in LSTM as one of externally controllable states, and extend the Hierarchical LSTMs to include mechanism for the integration. Experimental results demonstrate effectiveness of the proposed framework involving hierarchical LSTM and person-centered feature. In this study, we demonstrate improvement over the reference model [4] in two-stream LSTM based approach. Specifically, by incorporating the person-centered feature with meta-information (e.g., location data) in our proposed late fusion framework, we also demonstrate increased discriminability of action categories and enhanced robustness against fluctuation in the number of observed players.
引用
收藏
页码:155 / 163
页数:9
相关论文
共 40 条
[1]  
[Anonymous], 2015, ICLR
[2]  
[Anonymous], 2015, Recurrent neural network regularization
[3]  
[Anonymous], 2014, Generating sequences with recurrent neural networks
[4]  
[Anonymous], 2012, CRCV T 12 01
[5]  
[Anonymous], 2016, CVPR
[6]  
[Anonymous], 2014, NIPS
[7]  
[Anonymous], 2015, P 2015 C EMP METH NA, DOI [10.18653/v1/D15-1280, DOI 10.18653/V1/D15-1280]
[8]  
[Anonymous], 2015, P 3 INT C LEARN REPR
[9]  
[Anonymous], 2016, CONVOLUTIONAL 2 STRE
[10]  
[Anonymous], 2016, SOCIAL SCENE UNDERST