结合时序动态图和双流卷积网络的人体行为识别 
下载： 1095次
Human Action Recognition Combining Sequential Dynamic Images and Two-Stream Convolutional Network
中国民航大学天津市智能信号与图像处理重点实验室, 天津 300300
图 & 表
图 1. 动作表征整体流程示意图
Fig. 1. Overall flow diagram of action representation
下载图片 查看原文
图 2. 静态视频帧以及对应的时序动态图。(a)静态图像;(b)时序动态图;(c)光流图
Fig. 2. Static video frames and corresponding timing dynamic diagrams. (a) Static images; (b) timing dynamic diagrams; (c) optical flow diagrams
下载图片 查看原文
图 3. TS-CNN网络框架
Fig. 3. TS-CNN network framework
下载图片 查看原文
图 4. 不同子序列长度的识别结果
Fig. 4. Recognition results of different subsequence lengths
下载图片 查看原文
表 1不同输入方式下UCF101数据集识别准确率
Table1. Recognition accuracy of UCF101 dataset with different input modes unit: %
| Method | Split1 | Split2 | Split3 | Accuracy |
|---|
| SI | 84.6 | 84.9 | 85.0 | 84.8 | | SOF | 87.3 | 89.9 | 91.0 | 89.4 | | FSDI | 83.9 | 83.8 | 83.1 | 83.6 | | BSDI | 84.1 | 83.3 | 84.3 | 83.9 | | SDI | 85.7 | 86.2 | 85.5 | 85.8 | | ESDI | 87.2 | 86.8 | 87.6 | 87.2 | | SI+SOF | 93.2 | 94.0 | 94.2 | 93.8 | | ESDI+SOF | 94.8 | 94.6 | 95.3 | 94.9 |
|
查看原文
表 2不同输入方式下HMDB51数据集识别准确率
Table2. Recognition accuracy of HMDB51 dataset with different input modes unit: %
| Method | Split1 | Split2 | Split3 | Accuracy |
|---|
| SI | 54.8 | 50.4 | 49.6 | 51.6 | | SOF | 64.2 | 63.6 | 62.7 | 63.5 | | FSDI | 50.7 | 51.4 | 53.6 | 51.9 | | BSDI | 51.6 | 51.5 | 54.1 | 52.4 | | SDI | 54.5 | 52.9 | 53.7 | 53.7 | | ESDI | 53.6 | 55.5 | 55.6 | 54.9 | | SI+SOF | 68.7 | 67.5 | 68.4 | 68.2 | | ESDI+SOF | 69.6 | 71.2 | 71.6 | 70.8 |
|
查看原文
表 3不同融合方式在数据集中识别准确率
Table3. Recognition accuracy of different fusion methods on dataset unit: %
| Consensus function | UCF101 | HMDB51 |
|---|
| Max | 93.0 | 69.1 | | Average | 94.9 | 70.8 | | Weighted average | 93.8 | 69.7 |
|
查看原文
表 4不同网络模型在数据集中的识别准确率
Table4. Recognition accuracy of different network models on dataset unit: %
| Network structure | UCF101 | HMDB51 |
|---|
| Resnet101 | 93.6 | 68.4 | | Bn-inception | 94.2 | 68.2 | | InceptionV3 | 94.9 | 70.8 |
|
查看原文
表 5不同人体行为识别模型的识别准确率
Table5. Recognition accuracy of different human behavior recognition models unit: %
| Network | UCF101 | HMDB51 |
|---|
| Spatial stream | 84.8 | 51.4 | | Temproral stream | 89.4 | 63.5 | | Original two-stream | 88.0 | 59.4 | | Ref. [19] | 94.0 | 69.4 | | Appearance and long-sequential stream | 87.2 | 54.9 | | Short sequential stream | 89.9 | 64 | | TS-CNN | 94.9 | 70.8 |
|
查看原文
表 6不同算法的识别准确率
Table6. Recognition accuracy of different algorithms unit: %
| Feature extraction | Method | UCF101 | HMDB51 |
|---|
| Tradition | Ref. [7] | 84.8 | 57.2 | | Ref. [8] | 87.9 | 61.1 | | Deep learning | Ref. [17] | 88.0 | 59.4 | | Ref. [21] | 88.6 | -- | | Ref. [22] | 91.5 | 65.9 | | Ref. [23] | 93.1 | 63.3 | | Ref. [24] | 93.4 | 66.4 | | Ref. [19] | 94.0 | 69.4 | | Proposed | 94.9 | 70.8 |
|
查看原文
张文强, 王增强, 张良. 结合时序动态图和双流卷积网络的人体行为识别[J]. 激光与光电子学进展, 2021, 58(2): 0210007. Wenqiang Zhang, Zengqiang Wang, Liang Zhang. Human Action Recognition Combining Sequential Dynamic Images and Two-Stream Convolutional Network[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0210007.
PDF全文
