3D convolution with two-stream convNets for human action recognition

Karim Mohamed Hosny


Human action recognition is attempting to identify what kind of action is being performed in a given video by a person, it is considered one of the important topics in machine learning and computer vision. It’s importance comes from it’s need in many applications such as security applications and human computer interaction. Many methods have been researched to attempt to solve the problem, ranging from handcrafting techniques to deep neural network techniques and methods such as 3D convolution and recurrent neural networks has been used as well. Popular datasets have been curated in order to benchmark the methods researched to tackle this problem, datasets such as UCF-101 and HMDB-51 are the most popular and are being tested with for all current and past techniques in the area of human action recognition. two-stream convolutional networks, a deep learning technique, has picked up the trend in recent years to solve the human action recognition problem. Most famous method for solving the problem is by pre-processing the video to generate optical flow data or dense trajectories then feed them to a deep neural network alongside feeding static individual image frames of the video. We attempt to ask the question of can we classify human action without the need for pre-processing or handcrafted feature generation before using deep learning for classification? And how will 3D convolution affect the temporal stream and the overall classification accuracy. We contribute to solving the human action recognition problem by introducing a new end-to-end solution using two-stream convolutional network that learns static features and temporal features without any pre-processing for the data to generate optical flow or dense trajectories for video temporal information. Our method has been tested on UCF-101 and HMDB-51 datasets to compete with state of the art techniques. It shows that we were able to achieve high accuracy results without any pre-processing needed unlike current popular methods. Our method ranked among the highest in UCF-101, the only method which had a higher accuracy was a research modifying the original two-stream network by adding new fusion techniques. And ranked the highest in the HMDB-51 in comparison with the other techniques.