Human Action Recognition Based On Topic Model

Human action recognition is one of the basic tasks and the key technologies incomputer vision. It is widely used in intelligent video surveillance, human-computerinteraction, video conference, video retrieval and medical diagnosis. How to effectivelyextract the feature to action recognition, the knowledge in this area has theoreticalsignificance. This script focused on the problem of action description caused by thetarget motion and complex background. The main work of this paper is summarized asfollows:1) Interest points are often mistakenly detected because of complex and dynamicbackground. The accuracy of detection is affected by the wrong detection. This paperproposed a new algorithm based on saliency map and threshold matrix (SMTM-IPD) forinterest point detection. Firstly, the algorithm used saliency map to detect the rectangleregion around human. Then it used different thresholds for inside and outsider therectangle to reduce the number of interest points in background and effectively keep themore salient points around human. The experimental results on KTH and UCF databaseindicate that our algorithm can reduce the influence of background, highlight the objectand overcome the problem of single threshold. It shows that exclusion of irrelevantinterest points in background can reduce the false positive rate in action recognition.2) Due to the widely divergent action poses of human being, the interference ofnoise and various amplitudes of even the same action, we proposed a visual codebookgeneration method (3DSH) that merges several kinds of features. This method combinesthe features of3D-SIFT and HOOF. It maintains the merits of3D-SIFT asscale-invariance and noise-immunity, so the visual dictionary could describe the variousaction poses and the different amplitude of actions. The method also integrates HOOFfeatures for global action information to detect diverse action amplitudes. In comparedwith7other popular visual codebook generation method, our method increased by7.7% in average motion recognition rate. For the complex dynamics UCF motion dataset, themethod’s average recognition precision has increased by14%compared with other4methods.3)To address the problem of low precision which results from that existingrelationship of probability between different visual words in a video and a single visualword in different videos is not taken into consideration in parameter inference of LDAtopic model. A human action recognition system based on TMBP model is proposed.This method firstly represents LDA model as factor graph with belief propagation andeach word is arranged to every topic with a certain probability, and then preserve allposterior probability information. Video, visual word and action label correspond todocument, word and topic in original model, respectively. Thus the topic model cantotally apply to video processing. The experimental results indicate that introduction ofTMBP model to video processing can efficiently improve accuracy rate.4)To reduce precision loss of matching problem brought by non-rigid human objectand gesture variation while recognizing actions of multiple targets, we introduce analgorithm(KS-PE) based on Kalman filtering and human silhouette part feature. In ourmethod, we utilize Kalman filtering to predict the position of human objects and extractedge histograms from head, legs, and feet. By integrating the matching scores of thesehistograms, we can detect human objects instantly in the coming scene. Hence multiplehuman action recognition can be realized. We test our algorithm on UCF multi-objectdataset, the proposed algorithm shows that using human component silhouette feature wecan improve the recognition accuracy by3%.