Action identification using a descriptor with autonomous fragments in a multilevel prediction scheme

Recent technological advances have provided powerful devices with high processing and storage capabilities. Video cameras can be found in several different areas, such as banks, schools, stores, public streets and industry for a variety of tasks. Camera technology has increasingly improved, achieving higher resolution and acquisition frame rates. Nevertheless, most video analysis tasks are performed by human operators, whose performance may be affected by fatigue and stress. To address this problem, this work proposes and evaluates a method for action identification in videos through a new descriptor composed of autonomous fragments applied to a multilevel prediction scheme. The method is very fast and achieves over 90 % of accuracy in known public data sets. The developed system allows for the current video cameras the possibility of real-time action analysis, demonstrating to be a useful and powerful tool for surveillance purpose.The authors are thankful to FAPESP (Grants #2011/22749-8 and #2012/20738-1) and CNPq (Grant #307113/2012-4) for their financial support