Energy-Based Periodicity Mining With Deep Features for Action Repetition Counting in Unconstrained Videos

Action repetition counting is to estimate the occurrence times of repetitive motion in one action, which a relatively new, significant, but challenging problem. To solve this problem, we propose new method superior traditional ways two aspects, without preprocessing and applicable for arbitrary periodicity actions. Without preprocessing, proposed model makes our scheme convenient real applications; processing action more suitable actual circumstance. In terms methodology, firstly, extract features using ConvNets then use Principal Component Analysis algorithm generate intuitive periodic information from chaotic high-dimensional features; secondly, an energy-based adaptive feature mode selection adaptively select proper deep according background video; thirdly,we construct waveform based on high-energy rules by filtering irrelevant information. Finally, detect peaks obtain repetition. Our work two-fold: 1) We give significant insight that extracted recognition can well self-similarity action. 2) A mining rule presented, process actions preprocessing. Experimental results show achieves or comparable performance three benchmark datasets, i.e. YT_Segments, QUVA, RARV.