Action Recognition Using Frame Average Feature Map with 2D Convolutional Neural Network for Real-Time Video Analysis
Keon-Woo Kang, Hasil Park, Joongchol Shin, Jinsol Ha, Joonki Paik · 2020
A typical video action recognition system has a high computational cost and is not suitable for real-time applications. To solve this problem, we propose an action recognition method using a two-dimensional convolutional neural network (2D CNN), which has a significantly lower computational cost than a 3D CNN. In addition, the proposed method uses a small number of frames from the video for an accurate result. The proposed method consists of: i) pretrained VGG16,which is a2D CNN, to train the action of the video andii) a test with an average of ten frames per dataset. The proposed method improved the recognition performance with a reduce computational cost by using the average of several frames instead of directly analyzing all the frames for real-time video analysis environments.