Research On The Algorithm Of Group Activities Recognition In Video Based On Metric Learning

Group activities recognition is a frontier topic in the field of computer vision at present.It provides effective technical means for video surveillance in public security domain.With the rapid development of computer vision technology,technologies of image based object detection and recognition are becoming more and more mature.But the accuracy of the object detection and recognition algorithms based on video needs to be further improved.Group activities recognition is a kind of activities recognition.Because of the diversity of scenes,the different density of crowds,and the mutual occlusion between groups,it is much more complicated than single person behavior recognition.Although some achievements have been made in the research of this subject in recent years and there are some large databases available for research,most of the existing group activities recognition algorithms are complex,the recognition accuracy of the algorithms is not high,and the robustness is not strong.Distance metric plays a very important role in many applications in the fields of image understanding and pattern recognition.Euclidean metric is the most widely used.It considers the input sample space as isotropic,and it cannot fairly reflect the potential relationship between the dimension components of data samples.Mahalanobis metric and Cayley-Klein metric take account of the correlation between the dimension components of data samples,that is,the different dimension components of data samples are treated unequally,and they have better recognition performance than Euclidean metric in practical applications.This dissertation uses Mahalanobis metric and Cayley-Klein metric learning to realize the recognition of group activities in video to improve the accuracy and robustness of the recognition algorithms.The main research content and results are as follows:(1)An algorithm for group activities recognition based on M-DTC.WT and elliptical Mahalanobis metric is presented.Aiming at the problems of high algorithm complexity and poor recognition accuracy of algorithms for group activities recognition in video frames,the presented algorithm extracts features in frequency domain and classifies them based on distance metric.In order to increase the directional selectivity,M-DTCWT filters are constructed by adding the hourglass decomposition and reconstruction filters to DTCWT filters,which are used to decompose human images in video into multi-scale and multi-direction and obtain the high and low frequency coefficients.Features of high and low frequency coefficients are extracted.In order to better measure the similarity between unknown samples and reflect the potential relationship between the dimension components of data samples,elliptical Mahalanobis metric is used to classify the group activities features.The experimental results show that the proposed algorithm has a higher recognition accuracy than the classical classification algorithm such as support vector machine.(2)An algorithm for group activities recognition based on NSDTCWPT and Cayley-Klein metric learning is presented.NSDTCWPT has the characteristics of translation invariance and detail preservation.NSDTCWPT is used to decompose human body images in the video to get high and low frequency coefficients.In view of the deficiency of a single texture feature for the description of group activities,a feature description,the IDSC feature,is added on the basis of the ILBP feature.It can not only express the global information but also describe the local information,and it has better robustness to the body change and non rigid change of the target.In addition,the word bag model is introduced to effectively reduces the influence of the number of sampling points and location changes on group activities recognition.Cayley-Klein metric learning can model the potential relationships between samples and have a better distinction between group activities.The algorithm uses Cayley-Klein metric to learn and classify the fusion features,and completes the group acti-vities recognition in the video.Experimental results on Behave video set,Group Activity video set and self-built video set show that compared with support vector algorithm and Mahalanobis metric algorithm,the proposed algorithm has higher recognition accuracy.(3)A group activities recognition algorithm based on group structure features is presented.The above methods require accurate tracking for each person in the early stage,because the complexity of the scene and the existence of occlusion may cause tracking loss.To solve this problem,we use the location change of human joints to extract group structure characteristics,namely group center,motion histogram,intimacy histogram and central histogram.Then Cayley-Klein metric learning is used to learn the group structure features,and the group activities recognition in video is realized,The algorithm can better describe the relationship between human and human,and it is more robust and effective for the group activities recognition in complex scene.