Gait Recognition Under Occlusion Based On Deep Learning

Gait recognition,as a new biometrics recognition technology,has been widely studied by researchers in recent years.The gait recognition system realizes the identification of pedestrians by recognizing their unique posture during walking.Due to the characteristics of gait information such as uniqueness,easy to collect,non-contact,difficult to hide,difficult to camouflage,gait recognition has become one of the best methods for remote identification.However there are still many difficulties to be solved in the field of gait recognition.For example,in the real scene,pedestrians are easily blocked by obstacles or other pedestrians,which makes it difficult for the monitoring system to obtain the complete image of pedestrians.This will greatly reduce the recognition accuracy of existing gait recognition algorithms.To solve this problem,an image inpainting algorithm is proposed to eliminate the occlusion area in pedestrian images.In view of the insensitivity of Gait Energy Image(GEI)to single contour map,occlusion elimination image is proposed to overlay Reduction Gait Energy Image(R-GEI).And send the R-GEI into the gait recognition network to realize identity recognition.The main research contents of this article include the following:1.In this paper,based on the existing gait recognition dataset,the image processing method is adopted to synthesize the pedestrian gait recognition dataset under occlusion.Considering the real monitoring scene,this dataset includes two situations: pedestrians are blocked by obstacles and other pedestrians.2.In order to determine whether there is occlusion in the image and the location of the occlusion object,this paper proposes an occlusion detection algorithm.The algorithm is realized by similarity comparison between pedestrian contour image and GEI corresponding block.And a number of experiments are designed to verify the effectiveness of this algorithm.3.In the aspect of occlusion removal,this paper proposes an image repair method to eliminate the occlusion part in pedestrian images.In this paper,through principle analysis and experimental verification,the Edge-Connect algorithm is selected as the baseline model of the gait image occlusion removal algorithm.4.After determining the baseline model,this paper analyzes the shortcomings of the algorithm in the dataset,and proposes several improvement measures in the three aspects of network structure,attention mechanism and convolution mode.The most effective combination of improvement schemes is selected through experiments.The PSNR and MAE of the improved model are both greatly improved compared with the original model.5.In the aspect of gait recognition,the existing gait cycle detection methods are difficult to obtain the gait cycle under occlusion conditions.Therefore,NP-GEI,which is independent of gait cycle detection,is proposed in this paper.On the basis of this,R-GEI,which adopts the occlusion elimination contour,is proposed.The gait recognition under occlusion can be realized by feeding R-GEI into the gait recognition network.