Binary 'proximity patches motion' descriptor for action recognition in videos

Building action recognition systems that are simultaneously fast, robust and requiring small memory space is very challenging. Though current state-of-the-art frameworks proposed the best performance on accuracy, speed or memory, they do not offer simultaneously best performance on all three metrics. Thus, it is still possible to achieve a good trade-off among all these three metrics. Using a compact patch-based pattern, this paper introduces a novel binary motion descriptor to efficiently describe motion in video. The descriptor, namely the Proximity Patches Motion (PPM), compares in two different ways a 3 x 3 central patch centered on a detected keypoint with other 24 patches compactly positioned around it between three consecutive frames. Experimental results on the Weizmann and KTH datasets show that the proposed method not only requires a small amount of memory, but is also faster and achieves competitive accuracy when compared to the state-of-the-art spatio-temporal binary descriptors.