Pedestrian Re-identification Based On Improved Convolutional Neural Network

Pedestrian re-identification is a task for matching whether a pedestrian in a non-overlapping imaging area is the same pedestrian.In the field of intelligent security,it has broad application prospects.And the most typical application scenario is the “Skynet Action” and “Eye-Eye Project” that the country has vigorously carried out.Accroding to security system,we can use a query image to find if there is a matched pedestrian in gallery by computer.Because of the processing speed of computer,it can help find the correct one more quickly.Also it can greatly reduce manpower.However,due to factors such as illumination,posture,occlusion,viewing angle,and camera,the pedestrian images captured by the camera often vary widely,which makes the recognition difficult.For such complex recognition tasks,the traditional manual feature approach has been difficult.With the advancement of the convolutional neural network method in recent years,with its breakthrough in the field of signal processing,especially image recognition,many researchers have begun to apply convolutional neural networks to pedestrian re-identification,which is better than traditional features.The extraction method is better for recognition.Although the convolutional neural network has a better recognition effect than the traditional feature extraction method,there is still much room for improvement in the current recognition effect.In order to further improve the performance of pedestrian re-identification algorithm,this paper proposes a pedestrian re-identification model based on attitude orientation information to better fulfill the pedestrian re-identification task.At the same time,in view of the fact that the data of the current pedestrian re-identification data set is small,it is easy to make the model over-fitting.This paper adopts the GANs semi-supervised learning method for pedestrian re-identification and enhances ability of the model.The main content of this topic has the following two aspects:(1)According to the shortcomings of the current pedestrian re-identification method in the attitude orientation information method,a pedestrian re-identification model based on attitude orientation information is proposed.At present,the pedestrian recognition method only uses the information on the two-dimensional level of the picture to directly match,ignoring the influence of the angle of view on the matching,which will cause the pedestrian to be mismatched due to the difference of the rotating surface.The same is true for the best methods at the moment.In order to correct the mismatch matching and improve the recognition rate of pedestrian recognition,this paper proposes a pedestrian recognition model based on perspective orientation information.Based on the skeleton network,the perspective orientation prediction module is designed to reduce the misalignment caused by the angle of view.The depth separable module is introduced to prevent over-fitting of the model,and the extrusion and excitation module is introduced to reduce the interference of picture noise.The experiment proves that this paper re-identifies the effective recognition rate of pedestrian recognition based on the perspective of the pedestrian toward the information,and obtains better recognition effect than the past in the common public dataset.(2)For the problem that the current pedestrian re-identification data set is small and the model is over-fitting,this paper uses semi-supervised learning method for pedestrian re-identification.At present,based on the GANs semi-supervised learning method,it is difficult to train in the case of multiple classifications of discriminator models and the simple discriminator can not adapt to the problem of pedestrian re-identification.In order to solve the above problems,this paper proposes a split GANs semi-supervised learning method.The split GANs semi-supervised learning method separates the generation of the sample from the training,first trains a GANs model using a simple discriminator,extracts the generator in the GANs,uses the generator to generate unlabeled data,and then unlabeled data.The original training data is merged into a new training set,and the new training set is used for the training of the pedestrian re-recognition model.The unlabeled data is separately divided into one class for training,and finally the training is completed.Such a split GANs method does not require training a complex discriminator model to avoid the problem that the use of a complex discriminator results in a model that is difficult to train.The semi-supervised learning method of the separate GANs method effectively improves the generalization ability of the model and improves the performance of the pedestrian re-identification model.