The Implementation And Optimization Of A Deep Learning Based Action Recognition Algorithm

In the last few years,with the rapid development of mobile internet,and the appearance of customer-oriented modern mobile operating system,such as i OS and Android,etc.,combined with the effect of Moore's Law,it makes the intelligent mobile devices,e.g.smartphone,rapidly spread out and popular in our daily lives.With the precondition that smartphones are popularity and have user-friendly interface,those devices,with camera mostly,burst the production of video data.Besides,new ideas,internet of things,smart city,digitalized television channel,etc.,come into our lives and produce a huge amount of video data.As a result,the demand of video analysis raises up and becomes urgent.Action recognition in video sequences is an active research domain in computer vision,holds vast application scenarios in potential,such as in the field of surveillance,human-computer-interaction,healthcare,video retrieval,etc.In 2012,the breakthrough amazing result achieved by Alex Net in ImageNet competition indicates the effectiveness of deep learning and brings a revolution to the traditional computer vision research,especially to the spatial image analysis.In the meanwhile,the usage of deep learning in spatial-temporal video sequences is still in its early exploration stage.This paper mainly aims at the domain of action recognition in video sequences,designs and implements a deep-learning-based system for action recognition,would focus on its engineering implementation and optimization.It has three parts of main contents listed as follow:1.Based on deep learning algorithm,referred some of current papers and solutions,it overall describes a multi-stream algorithm for action recognition in spatial-temporal video sequences.This multi-stream approach has five analysis streams,still frame stream,optical flow stream,shot clip stream,long clip stream and audio stream,etc.2.Introduces the framework design of this action recognition system in overall,the basic features of each main component and key data structures in detailed.Then it presents two core sub-systems,training system and inference system in a few sections and the common components they shared,also describes the performance optimization for some components.3.Tests the whole system and algorithm,analyzes the key statistics,such accuracy,calling cost,false alarms,etc.and compares it with some other approaches.In the end,the test results show that the algorithm and system could achieve a brilliantperformance in accuracy with a reasonable cost.