Hicle Detection And Attribute Recognition Based On Deep Learning

Vehicle detection and attribute recognition is a fundamental task in Intelligent Transport System(ITS)which contains many basic functions such as vehicle detection,vehicle color recognition,vehicle type recognition,vehicle brand recognition and so on.Most of current solutions are task-oriented that these methods focus on accomplishing vehicle detection and attribute recognition subtasks one by one with off-the-shelf deep learning algorithms,ignoring the properties of different subtasks and the relations between them.Such kind of solutions make the vehicle detection and attribute recognition system complicated and lower computational efficiency.To address these issues,this dissertation propose a novel vehicle detection and attribute recognition framework which only contains two main modules,the first one is course attribute recognition and detection module and the second one is fine-grained attribute recognition module.Firstly,we bring out a unified vehicle detection and attribute recognition algorithm for course attribute recognition and detection module which integrates vehicle color and type information into vehicle detection algorithm.The proposed framework utilize the ability of multi-task learning algorithm to model the vehicle color recognition,vehicle type recognition and vehicle detection tasks at the same time.In order to handle the long-tail distribution problem,the Online Hard Example Mining(OHEM)algorithm is added into the training pipeline.Secondly,we design a vehicle brand recognition algorithm for fine-grained attribute recognition module,solving the vehicle brand recognition problem with more compact convolutional neural networks architecture and more discriminative loss function.The vehicle brand recognition architecture is built with high performance microarchitecture in GoogleNet and ResNet.In order to learn more discriminative features,center loss and cross-entropy loss are both introduced into the vehicle brand recognition model.Furthermore,we apply active learning paradigm into the vehicle brand recognition algorithm which enables the model with the ability of recognizing unknown-class example.In order to verify the performance of proposed vehicle detection and attribute recognition algorithms,we construct two large-scale dataset which are on-road vehicle image dataset and vehicle brand image dataset.The unified vehicle detection and attribute recognition algorithm outperform SSD and Faster-RCNN detection algorithm on detection task.The vehicle brand recognition algorithm not only achieve high recognition accuracy for known-class examples but also reach high recall performance for unknown-class example.In summary,with the proposed framework in this dissertation we solve the drawbacks existed in the task-oriented solutions and produce more powerful vehicle detection and attribute recognition performance while lower the computational cost.