Research On Gait Recognition Method Based On Pedestrian Global And Local Features

Gait recognition is a method of identifying pedestrians by their walking postures.As an emerging biometric technology,compared with other biometric technologies,gait recognition has the unique advantages of being easy to capture at long distance,difficult to camouflage,non-contact and no cooperation.It can play an important role in public safety,which has attracted more and more research attention in recent years.However,gait recognition technology is immature at present,and there are still many challenges,such as the variations of viewing angles,pedestrians’ backpacks,clothing items and background interference,and pedestrian occlusions,which all bring many difficulties to gait recognition.In this paper,the gait recognition technology is studied using the method of deep learning,and we propose a gait recognition method based on the global and local feature of pedestrians.The main work is as follows:First,by observing the gait patterns of pedestrians,it is found that different body parts of pedestrians have different changes and motion patterns during walking.For example,the head of the human body changes slightly during walking,while the legs and arms sway to a large extent,and show a certain style,so the extraction of these local features is more conducive to gait recognition.Therefore,when extracting frame-level gait features,this paper proposes to extract both global and local frame-level features from a set of gait silhouette of pedestrians.Second,in order to better preserve temporal and spatial information,the extracted framelevel features need to be aggregated into set-level features.Since the frame-level feature extraction process includes multiple stages of extraction,the obtained global frame-level and local frame-level features are further aggregated into the global set-level features and local setlevel features through the improved Set Pooling operation.Third,the final set-level features and frame-level features are used with horizontal pyramid module and temporal feature aggregator module,respectively.The horizontal pyramid can horizontally divide the set-level feature into multiple horizontal block spaces at multiple scales,and horizontal pyramid pooling is used for each horizontal block space to obtain more detailed local features;Temporal feature aggregator can aggregate locally adjacent frame-level features to obtain gait features with short-term features.Finally,the feature produced by the horizontal pyramid and temporal feature aggregator module is passed through a fully connected layer to increase the discriminative ability of the feature for the final recognition.Finally,this work uses the CASIA-B and CASIA-C gait data set to verify the feasibility of the algorithm,and obtains a good recognition effect,which reaches 96% in the normal walking state,92.0% in the backpack state,and 76.5% in the clothing state;in CASIA-C it reaches an average of 97.5%.