Research On Motion Detection Method Based On Multi-instance Learning

Temporal action localization in untrimmed long videos is a key technology in the fields of intelligent monitoring and video analysis.The task of action localization is expected to output not only the action category,but also the precise start time and end time.In recent years,with the development of computer vision and deep learning theory,many breakthroughs have been made in action localization task.However,there are still some problems such as imprecision of detected action boundaries and inaccuracy of classification.Therefore,this paper proposes a temporal action localization method and an action proposal generation method based on multiple instance learning.In the last we implement a real-time action localization system.The detailed work of this paper is as follows:Firstly,we propose a novel framework which firstly models each action clip based on its temporal evolution,and then adopts a deep multiple instance learning(MIL)network for jointly classifying action clips and refining their temporal boundaries,which can address precise temporal action localization.Secondly,we propose an action proposal generation method which can retrieving proposals to cover truth action instances with high recall and high overlap.Likewise,we adopt a deep multiple instance learning(MIL)network to estimate actionness scores and refine boundaries simultaneously,which is trained via a multi-task loss with smoothness constraint and coordinate regression.During post-processing step,we use soft non-maximum suppression to suppress redundant proposals more effectively.Last but not least,we design and implement an action localization system based on our proposed methods.It is convenient and intuitive to observe the localization results and the localization log of action localization methods through this system.Moreover,we separate the display module and algorithm processing module which makes this system can be reused by other action localization methods...