Temporal Action Detection Based On Deep Learning

In recent years,with the global mobile Internet technology and the rapid development and popularization of digital equipment,video data on the Internet growing at an alarming rate every day,rely on artificial way of dealing with such a massive video data is not reality,based on the deep learning of temporal action detection research aroused the interest of the researchers,has become the research hotspot in the field of intelligent video analysis.The research involves computer vision,image processing,video analysis,target detection and many other research fields,has important scientific significance,can be widely used in intelligent security,robot vision,virtual reality,video surveillance,human-computer interaction and other fields,has a good application prospect.Temporal action detection is to detect the action fragments in a given uncropped video,including the start time,end time and action category.It is one of the research hotspots in the field of intelligent video analysis.Traditional methods based on manual feature extraction have poor robustness for complex and varied motion types,while deep learning methods can effectively learn the differences between different movements.A large number of research achievements have been made in the field of video motion analysis.But temporal action detection performance badly depends on the action of the candidate proposed effect,effective candidate proposal played a decisive role in action detection effect,and video data structure,diverse and complex,target movement change action duration varying lengths,makes the temporal sequence in the action detection proposed method there is video features using such problems as inadequate,the target action boundary detection difficult.Aiming at the problems above,this paper introduced the expansion convolution,branch expansion was proposed based on multiple convolution of temporal action detection algorithm,through the different expansion rate of convolution operation to detect multi-scale action,effectively improves the recall rate of the segments of different scales action,on this basis to further introduce attention mechanism,automatic action of the importance of learning different scales,enhanced the sensitivity of the model of action boundaries,and put forward the evaluation model,based on the time convolution networks through a more accurate assessment score,improve the positioning accuracy of the boundary.In order to verify the effectiveness of the method proposed in this paper,an experimental comparison was made with the most recent relevant algorithms on the open Activity Net and THUMOS-14 datasets.The experimental results show that the proposed method has a good effect on the detection performance and has achieved a certain improvement.