时序行为提名的上下文信息融合方法

doi:10.3778/j.issn.1673-9418.1912044

摘要/Abstract

摘要：

在针对视频的人体活动定位和识别领域中，现有的时序行为提名方法无法很好地解决行为特征长期依赖性而导致提名召回率较低。针对此问题，提出了一种上下文信息融合的时序行为提名方法。该方法首先采用三维卷积网络提取视频单元的时空特征，然后采用双向门控循环网络构建上下文关系预测出时序行为区间。针对门控循环单元（GRU）存在参数较多和梯度消失的问题，通过输入特征控制门结构增强并行计算能力，通过引入加权平均增强历史和当前时刻信息融合能力，提出了一个简化的门控循环单元（S-GRU）。最后在数据集Thumos14上进行实验验证和比较，结果表明基于双向S-GRU循环网络的时序行为提名方法提高了提名召回率。

关键词: 门控循环网络（GRU）, 梯度消失, 上下文信息, 时序行为提名, 时序行为检测

Abstract:

In the field of human activity localization and recognition in videos, the existing temporal action proposal methods have not solved the long-term dependence problem better, which results in lower recall rates of proposals. In view of this problem, a method based on context information fusion for temporal action proposals is proposed in this paper. Firstly, the spatiotemporal features of video units are extracted by the 3D convolutional network. Then, the bidirectional recurrent network is used to construct the context relationship for predicting the temporal action proposals. Considering the problems of more parameters and the vanishing gradient in the gated recurrent unit (GRU), a simplified-GRU (S-GRU) is proposed, in which the input features control the gating structure to enhance the parallel computing capability and the weighted average is introduced to enhance the ability of the gated recurrent unit to adaptively fuse the history and current time information. Finally, experimental results on the Thumos14 dataset demonstrate that the method based on the bidirectional S-GRU for temporal action proposals improves the recall rate of proposals.

Key words: gated recurrent network (GRU), vanishing gradient, context information, temporal action proposals, temporal action detection

王新文, 谢林柏, 彭力. 时序行为提名的上下文信息融合方法[J]. 计算机科学与探索, 2021, 15(3): 486-494.

WANG Xinwen, XIE Linbo, PENG Li. Context Information Fusion Method for Temporal Action Proposals[J]. Journal of Frontiers of Computer Science and Technology, 2021, 15(3): 486-494.

参考文献

[1] SHOU Z, WANG D G, CHANG S F. Temporal action locali-zation in untrimmed videos via multistage CNNs[C]//Pro-ceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, Jun 27-30, 2016. Wash-ington: IEEE Computer Society, 2016: 1049-1058.
[2] DUCHENNE O, LAPTEV I, SIVIC J, et al. Automatic annot-ation of human actions in video[C]//Proceedings of the 2009 IEEE International Conference on Computer Vision, Kyoto, Sep 29-Oct 2, 2009. Piscataway: IEEE, 2009: 1491-1498.
[3] FENG X Y, MEI W, HU D S. Aerial target detection based on improved faster R-CNN[J]. Acta Optica Sinica, 2018, 38(6): 242-250.
冯小雨, 梅卫, 胡大帅. 基于改进Faster R-CNN的空中目标检测[J]. 光学学报, 2018, 38(6): 242-250.
[4] REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceed-ings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, Jun 27-30, 2016. Washington: IEEE Computer Society, 2016: 779-788.
[5] ESCORCIA V, HEILBRON F C, NIEBLES J C, et al. DAPs: deep action proposals for action understanding[C]//LNCS 9907: Proceedings of the European Conference on Computer Vision, Amsterdam, Oct 11-14, 2016. Berlin, Heidelberg: Springer, 2016: 768-784.
[6] BUCH S, VICTOR E, ESCORCIA V, et al. SST: single-stream temporal action proposals[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Jul 22-25, 2017. Washington: IEEE Computer Society, 2017: 6373-6382.
[7] GUO D S, LI W, FANG X Z. Fully convolutional network for multiscale temporal action proposals[J]. IEEE Transac-tions on Multimedia, 2018, 20(12): 3428-3438.
[8] GAO J Y, YANG Z H, CHEN S, et al. TURN TAP: temporal unit regression network for temporal action proposals[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 24-27, 2017. Piscataway: IEEE, 2017: 3648-3656.
[9] TRAN D, BOURDEV L, FERGUS R, et al. Learning spatio-temporal features with 3D convolutional networks[C]//Pro-ceedings of 2015 IEEE International Conference on Computer Vision, Santiago, Dec 13-16, 2015. Piscataway: IEEE, 2015: 4489-4497.
[10] CARREIRA J, ZISSERMAN A. Quo Vadis, action reco-gnition? A new model and the kinetics dataset[C]//Proceed-ings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 24-27, 2017. Piscataway: IEEE, 2017: 4724-4733.
[11] LI Q H, LI A H, WANG T, et al. Double-stream convolutional networks with sequential optical flow image for action rec-ognition[J]. Acta Optica Sinica, 2018, 38(6): 226-232.
李庆辉, 李艾华, 王涛, 等. 结合有序光流图和双流卷积网络的行为识别[J]. 光学学报, 2018, 38(6): 226-232.
[12] SONG L F, WENG L G, WANG L F, et al. Multi-scale 3D convolution fusion two-stream networks for action recogni-tion[J]. Journal of Computer-Aided Design & Computer Graphics, 2018, 30(11): 99-108.
宋立飞, 翁理国, 汪凌峰, 等. 多尺度输入3D卷积融合双流模型的行为识别方法[J]. 计算机辅助设计与图形学学报, 2018, 30(11): 99-108.
[13] SCHUSTER M, PALIWAL K K. Bidirectional recurrent neural networks[J]. IEEE Transactions on Signal Process-ing, 1997, 45(11): 2673-2681.
[14] PASCANU R, MIKOLOV T, BENGIO Y. On the difficulty of training recurrent neural networks[J]. arXiv:1211.5063, 2012.
[15] HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780.
[16] CHO K, VAN MERRIENBOER B, GULCEHRE C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Pro-cessing, Qatar, Oct 25-29, 2014. Stroudsburg: ACL, 2014: 1724-1734.
[17] HOSANG J, BENENSON R, SCHIELE B. A convnet for non-maximum suppression[C]//LNCS 9796: Proceedings of the 38th German Conference on Pattern Recognition, Han-nover, Sep 12-15, 2015. Berlin, Heidelberg: Springer, 2016: 192-204.
[18] JIANG Y, LIU J, ROSHAN Z A, et al. Thumos challenge: action recognition with a large number of classes[EB/OL]. [2019-11-20]. http://crcv.ucf.edu/THUMOS14/.
[19] SOOMRO K, ZAMIR A R, SHAH M. UCF101: a dataset of 101 human action classes from videos in the wild[EB/OL]. [2019-11-20]. http://crcv.ucf.edu/data/UCF101.php.
[20] LI N N, GUO H W, ZHAO Y, et al. Active temporal action detection in untrimmed videos via deep reinforcement learn-ing[J]. IEEE Access, 2018, 6: 59126-59140.