Four-channel behavior recognition algorithm based on DRN
Nan Yang, Shen Yang, Jin Wu, Huiping Deng · 2018
Concerning the fact that previous behavior recognition algorithms tend to be affected by background and illumination noise, we propose a four-channel behavior recognition algorithm based on fine-tuning deep network Dilated Residual (DRN). Our algorithm can extract more human behavior information through the high frequency grayscale channel. Specifically, we decompose video datasets into multiframe color images, and convert them to high frequency grayscale images with high-pass filtering and normalization. Then, the four-channel dataset consists of RGB and above high frequency grayscale images. The effectiveness of the algorithm is evaluated by testing the classification accuracy on Olympic Sports dataset with DRN. The result shows that the algorithm achieves 85.5% accuracy on behavior recognition, which is 3.4% higher than the TOI+SVM method.