The Research On Pedestrian Recognition Based On Biological Vision Model

Video surveillance has found a wide range of applications in areas such as intelligent transportation,individual and group behavior identification and analysis,security monitoring and criminal suspect tracking.Owing to large scene size,complex background and poor lighting conditions,it is challenging to extract desired information such as face features.Currently manpower inspection is the main method for searching videos for a person.Such an operation mode is inefficient and prone to detection error.Therefore,it is of great significance to apply intelligent science to image processing so as to achieve automatic search.This thesis develops a new recognition model based on two typical biological vision models: the spiking neural network and the convolutional neural network.Following the information exchange mechanism in the spiking neural network model,the recognition model filters out non-critical information by means of adjusting the threshold of the spiking neurons.On the other hand,the new model automatically extracts the critical features of the dataset,adopting the receptive field model and the layering principle in the convolutional neural network model.Both the filtering and the extracting are implemented using rapidly developing GPU parallel computing technology.In the field of pedestrian recognition,one of the most popular applications is to extract features based on a person's photos and search for the person in videos recorded by different cameras.The pedestrian search can be achieved by means of a trained classifier,when multiple pedestrian photos are available for training set,and the person re-identification,when pedestrian photos are lacking.This thesis studies both pedestrian classification and person re-identification.For the classification issue,this thesis proposes a model based on spike neural network that extracts gabor texture features using the simple cell receptive field model used in biological vision system and the basic principle of spike neural network.The texture features extracted by spike neural network were tested on the zero-angle images of the gait dataset CASIA Dataset A,and the recognition rate is higher than that using the gabor features.This thesis also develops a multi-channel pedestrian recognition algorithmthat is adopted from the biological vision model and searches pedestrians by means of color,texture and angle histogram features.Testing of the algorithm was conducted on a specially generated multi-camera videos dataset that captures different pedestrians in different scenes under different lighting conditions and generated multiple results that validate the algorithm.For pedestrian recognition,this thesis proposes a deep pedestrian multi-part and multi-kernel metric network that consists of deep learning model and metric learning.While the deep learning model learns the local and global features of pedestrians,the metric learning spatially transforms pedestrians' features and reduces the distance of the same pedestrian in the new space.Compared with the experimental results of the multi-part triplet model with improved loss function and the multi-kernel canonical correlation analysis based on artificial features,the top-1 accuracy rate of the network model is 2.96% and 2.91% higher on VIPeR,and 8.1% and 3.4% higher on PRID.The network is highly scalable and flexible,can be adjusted with a new deep learning model according to a specific problem.Or it can use different kernel functions or other metric mechanisms.Based on the deep pedestrian multi-part and multi-kernel metric network model,this thesis improves the deep learning model: it designs a deep multi-part stn-triplet network,adds the spatial transformer network to the deep learning model,simplifies the model branch,adds a sampling layer,and uses the sampling strategy so as to reduce the input combination of the loss layer and accelerate network training.The top-1 accuracy rate has increased by 3.07% and 0.48% on VIPeR and GRID,respectively.This thesis also presents a deep multi-attribute quadruplet network that replaces the triplet network in the deep pedestrian multi-part multi-kernel metric network.The four-input quadruplet network is designed to extract pedestrian attribute features from VIPeR and GRID datasets and reduce the distance between the same attribute of the same pedestrian.Compared with the attribute classification method,the quadruplet network increases top-1 accuracy rate by 28.52% and 5.28% on VIPeR and GRID,respectively.