Combining gradient histograms using orientation tensors for human action recognition

We present a method for human action recognition based on the combination of Histograms of Gradients into orientation tensors. It uses only information from HOG3D: no features or points of interest are extracted. The resulting raw histograms obtained per frame are combined into an orientation tensor, making it a simple, fast to compute and effective global descriptor. The addition of new videos and/or new action cathegories does not require any recomputation or changes to the previously computed descriptors. Our method reaches 92.01% of recognition rate with KTH, comparable to the best local approaches. For the Hollywood2 dataset, our recognition rate is lower than local approaches but is fairly competitive, suitable when the dataset is frequently updated or the time response is a major application issue.