Vehicle Re-identification In Surveillance Videos

Video surveillance,as one of the famous topics in the field of computer vision and pattern recognition,is widely used in the public security,intelligent transportation and etc.Vehicle re-identification is one key technology in video surveillance,which could be used to suspect vehicle tracking and trajectory prediction.Vehicle re-identification is a retrieval process of complex images and is to find the target vehicle in the videos taken by non-overlapping cameras at different time.In order to realize vehicle re-identification,moving vehicle detection and feature extraction algorithm are studied,and a vehicle re-identification system with good accuracy is devised.The main work is summarized as follows:First,the moving vehicle detection algorithm is studied and implemented.In order to accelerate the speed of vehicle detection,ViBe background modeling algorithm is adopted to obtain candidate regions of vehicle.On the basis of motion detection,deformable part model is studied and analyzed,which could be used to the detection of multi-pose vehicle,so the vehicle tracking sequences are obtained.Then,as for the feature extraction,the color feature descriptor is proposed firstly,which can keep balance between computational complexity and performance.To extract the color descriptor,based on the nonlinear partition of HSV color space,the color look-up table is established,and the high dimensional color space is mapped to low dimension.In addition,since deep learning has achieved great success in face recognition,image classification,object detection and so on,deep learning is studied in this paper.Firstly,AlexNet is studied,by fine-tuning the pre-trained AlexNet and training it using vehicle dataset,the feature could be extracted by the trained model.Secondly,a Siamese network for vehicle verification is designed.Finally,an improved Siamese neural network for vehicle re-identification is proposed,the network's loss function is composed of Contrastive loss and Softmax loss,which can reduce the intra class distance and increase the distance between classes,the network could improve the discriminability of feature.Finally,this paper uses the color feature by manual design and deep feature by convolutional network for vehicle re-identification,the experiment shows that compared with the low-level features,the deep learning can extract some distinguished features.And the improved Siamese network proposed by this paper gets the best result.