The Study Of Fast Vehicle Re-identification Using Deep Learning

Intelligent Transport System(ITS)is one of the key department in metropolitan cities to ensure the smooth and safe mobility of vehicles.The vehicle re-identification(Re-Id)is among an essential task,defined as the problem of identifying vehicle across images that have been captured by different surveillance cameras without overlapping fields of view.In other words,either any specific vehicle captured in one camera has already appeared over multiple camera network or not.With the increasing need for automated video analysis,the defined task is receiving increasing attention these days.It underpins many critical applications such as intelligent vehicle parking,suspicious vehicle tracking,vehicle incident detection,cross camera tracking,road access restriction management system,and automatic toll collection.In the past few years,various powerful computer vision approaches are exploited to deal with such video analysis of surveillance cameras for vehicle Re-Id task.However,due to unique task oriented requirements,it became more challenging for the researchers to design robust and efficient model to address related issues such as inter-class similarity,viewpoint changes,partial occlusions,intra-class variability,background clutter,and cross dataset vehicle Re-Id;previously proposed models are not much effective to deal with mentioned problems.In this dissertation,we aim at exploring the way to tackle different problems and propose deep learning based techniques to obtain better vehicle Re-Id performance in different ways.Initially,we propose a fast model for traffic surveillance to identify different types of vehicles found in the camera network.We adopt deep convolutional neural network(CNN)model Inception-v3 for visual features vector extraction,due to transfer learning,model accelerate the learning process and minimize the training time.In order to make use of features vector for classification between each class,three different algorithms are implemented and investigated to distinguish vehicle types(small,medium,and large).After differentiating type of vehicles,the salient features extraction and selection are crucial aspects for the vehicle Re-Id task.To solve this problem,as our second contribution focuses on informative features extraction and selection by using global region channel and the local region channel.The global region channel extracts the feature vector from the whole vehicle image,whereas,the local region channel extracts more discriminative and salient features from different patches of the same image.In addition to this,we incorporate attributes like model,type,and color of vehicle.Our proposed approach re-identifies the vehicle in two steps: first,shortlist the vehicle from a gallery set based on appearance,and then the next step verifies the shortlisted vehicle’s license plates with a query image to identify the targeted vehicle.Previous approaches for vehicle Re-Id task do not perform well if dealt with biased vehicle Re-Id datasets.The datasets should represent the real-world scenario for the practical implementation of the model.However,it has been noticed that most of available datasets are biased and co-related,which leads towards the highly over-fitted model.It has become a non-trivial task to deal with such kind of biased datasets.As the third contribution,to investigate the vehicle Re-Id datasets bias problem,we first create the aggregated dataset that consists of different vehicle Re-Id datasets,where each image in this new dataset is labelled with the same name as parent dataset name.Afterwards,we classify the Re-Id datasets using deep CNN model Inception-v3 by fine-tuning the last layer.Dataset classification results also verify our hypothesis and indicate that currently available datasets are highly biased.To mitigate this issue,we present a novel data augmentation technique by inserting an additional type of variability in the training set,without any computationally expensive procedure.In the last,we observed that most vehicle Re-Id approaches are evaluated on a single dataset in which the training and testing of the model is performed on the same dataset.However,this negatively affects model generalization ability due to biased datasets along with the significant difference between training and testing data.Hence,the model becomes impractical in a real-world environment.As the last contribution,to demonstrate the effect of dataset bias issue,we conducted a deep investigative study of existing approaches on cross dataset vehicle Re-Id.The conducted study supports our argument that the performance of existing approaches is compromised over cross dataset vehicle Re-Id.In this regard,we propose an approach by utilizing spatio-temporal information and transfer learning technique with appearance to find out exact vehicle fast;however,it very hard to find out exact matching very fast using appearance features.In this approach,we augment the data during training,whereas,spatio-temporal patterns of unlabelled target dataset are learned by transferring siamese neural network(SNN)classifier trained on source labelled dataset.We finally calculated the composite similarity score of spatio-temporal pattern with SNN classifier visual features.All the proposed approaches in this dissertation are evaluated on publicly available benchmark datasets and show promising results towards the effectiveness on this problem as compared to existing state-of-the-art existing approaches.