Human Activity Analysis In Videos

In recent years, with the development of video capture technology, more andmore people prefer to use cameras to record their daily life. Thus, the amount ofvideos is growing explosively. It has been an important topic in the field ofcomputer vision to analyze the video content by machine learning ways. As most ofvideos record human-centered content, human activity analysis has become the mostactive subject among many branches of content-based video analysis, with a widerange of applications, including intelligent human-computer interaction, videosurveillance, telemonitoring of patients, etc.In this paper, we focus on human-centered videos and firstly propose to analyzehuman activities in a supervised learning manner. By this way, we can classify thesingle person actions which have lower intra-class divergence and are much easierto be labeled and modeled. Supervised learning, however, can’t be used for all kindsof human activities, such as crowded activities with large intra-class divergence.Therefore, we then propose an unsupervised analysis method to discover meaningfulcrowded activity patterns in the data.In order to achieve the supervised analysis of single person actions, we proposea spatial-temporal constraint-based action recognition method, in which two actionsare compared by both the appearance features and the spatial temporal structures.We first utilize a random quantization method to obtain more comprehensive andaccurate matching of appearance feature points by combining multiple quantizationresults from different perspectives. Then, we map the matched pairs of points to aspace-time offset space to indirectly gain the matching of spatial-temporal structuresshared in feature points, thus we get more accurate similarities between actions.Finally, we leverage the KNN model to complete the classification of the action. Theexperiment results on both KTH action dataset and YouTube action datasetdemonstrate the effectiveness of our proposed action classification method.In the analysis of crowded activities, we develop a novel unsuperviseddiscovery algorithm aiming to automatically explore latent action patterns amongcrowd activities and partition them into meaningful clusters. Inspired bycomputational model of human vision system, we present a spatiotemporalsaliency-based representation to simulate visual attention mechanism and encodehuman-focused components in an activity stream. Combining with feature pooling,we could obtain a more compact and robust activity representation. Based on theaffinity matrix of activities, N-cut is performed to generate clusters with meaningful activity patterns. We carry out experiments on our proposed HIT-BJUT dataset andanother public UMN dataset. The experimental results demonstrate that theproposed unsupervised discovery method is capable of automatically miningmeaningful activities from large scale video data with mixed crowd activities.