Study On The Key Technology For Human Activity Recognition

Human activity recognition has become one of the most active research topics inthe artificial intelligence and pattern recognition field recently, due to its wideapplications. The key technologies for human activity recognition, including the imagepreprocessing, activity representation, feature reduction, as well as the activityclassification technology are all thoroughly studied in this paper. Based on them, wehave put forward several novel activity recognition methods applicable to differentsystems, and achieved satisfactory results. The main contribution of this thesis can beconcluded as follows.(1) As a rule of thumb, the effective image preprocessing is crucial to the ultimaterecognition performance of human activity. We put forward a color fusion algorithm forthe dual-band surveillance application. The visible and the thermal infrared sourceimages are fused with the non-subsampled contourlet transform (NSCT), andsubsequently colorized with the color transfer scheme. Experimental resultsdemonstrate that the method can not only keep abundant details of the background, butalso improve the human target detectability. As a result, the situation awareness can beenhanced, and new source images can be obtained with higher quality. Moreover,experiments also show that the fused images can improve the robustness and accuracyduring the target tracking process.(2) We put forward a novel human activity recognition method based on a newspatial-temporal template called variation energy image (VEI), which can betterrepresent the shape and the motion features. The multi-class relevance vector machines(mRVM) is introduced into the human recognition field for the first time, which is thecurrent state-of-the-art kernel machine learning technology given the multi-classclassification problems. We have achieved a recognition rate as high as98.2%on theWeizmann dataset, which prove that the mRVM especially the mRVM2has advantagesboth in terms of recognition rate and sparsity. Moreover, we further found that ourrecognition rate is higher than other methods, mainly resulting from the differences inactivity representation and activity classification process.(3) We also put forward several human activity recognition methods based on thehuman-vision properties, and test them on a thermal infrared human activity datasetconstructed by Chongqing University. Both the Gabor and Log-Gabor wavelets are employed to describe the infrared human activity for the first time, of which we stronglyrecommend the latter so as to reduce the required scale number. The principlecomponent analysis method and the discriminative common vectors method are used tosolve the feature reduction problem, as well as the small sample size problem,respectively. A recognition rate as high as94.44%is achieved on the infrared dataset,and the influence of several factors (including the Gabor-based wavelets, featurereduction method as well as the classification method) on the recognition performanceis thoroughly studied.(4) Finally, we studied the human activity recognition with wearable sensor. Weput forward a novel method in which the generalized discriminant analysis is used forthe first time to reduce the high dimensional time and frequency features, derived fromthe multiple sensors. And then an array of relevance vector machines is constructed toclassify the reduced features. Experimental results on the WARD dataset with differentclassification techniques demonstrate that our approach can achieve the best recognitionrate as high as99.2%. In consideration of the robustness as well the optimizablemodifications for the multi-sensor system, we further studied on the decision fusionmethods. The relation between several factors (including the number of sensors as wellas the deployment way, the fusion rule, the feature reduction method and theclassification method) of the recognition method and the recognition performance arethoroughly investigated.