Multi-Gait Recognition And Synthesis

Gait represents a person's walking style.Gait recognition has attracted more and more attention in pattern recognition field since it can recognize a subject from a long distance and is difficult to disguise.Currently,there are mainly three kinds of influencing factors in gait recognition: dress,carrying and view.To address the influence of these three influencing factors,existing methods can be classified into two types:model-based methods and appearance-based methods.Existing methods have achieved satisfactory performance in gait recognition of single person walking.However,in realities,gait patterns include not only single-person walking but also multiple-person walking.We define the walking patterns of multiply persons walking together as multigait.In condition of multiple persons walking,each person's gait produces changes to a certain extent.These changes are different between different participants.Thus,traditional gait recognition methods cannot achieve good performance in multi-gait recognition.To solve this problem,this paper studies the changing regularity of multi-gait,and how to complete multi-gait recognition based on the changing regularities.The main research contents include:1.This paper concludes the typical frameworks of existing single-gait recognition methods,and analyzes the challenges of these methods in multi-gait recognition.2.To solve recognition problem caused by gait variance in multiply person walking,this paper proposes an attribute discovery model based on latent conditional random field(CRF)model,which aims to discover automatically stable gait attributes.The discovered stable attributes are used to complete multi-gait recognition.The experimental results suggest that recognition accuracy the proposed method outperforms existing gait recognition methods significantly.However,when gait changes a lot from single-gait to multi-gait,the number of discovered stable attributes may be too small to reflect adequate gait features.3.To study effect of traditional gait recognition methods on multi-gait recognition,this paper designs a multi-gait segmentation method based on hypergraph parti-tion.This method first segments images into grids,then constructs a hypergraph based on the information like shape,color and so on in each grid.Then different participants are segmented by hypergraph segmentation.After gait images of each person are segmented,traditional single-gait recognition methods can be applied in multi-gait recognition.This method can preserve unoccluded gait regions as much as possible.4.This paper studies gait segmentation methods based on human detection and tracking technologies.Since multi-gait suffers from continuous irregular occlusion,human contours achieved by human detection and tracking technology are often incomplete.To address the problem,this paper proposes a high-dimensional exemplar-based method to inpaint original gait images.We also inpaint two-value image sequences by low-rank tensor.The inpainted gait images are represented by tensors.The experimental results suggest that the proposed method achieves high multi-gait recognition accuracy.5.This paper proposes multi-gait synthesis method to realize transformation between single-gait and multi-gait.The synthesis is realized based on sparse coding.We first encode image sequence of single person sparsely,and then predict sparse codes of multi-gait by kernel regression.Synthetic multi-gait image sequences can be obtained by decoding process.To evaluate multi-gait synthesis effect quantitively,we apply synthetic multi-gait images to multi-gait recognition and use the recognition accuracy as the metric.