Research On Vehicle Type Recognition And Detection Algorithm In Surveillance Scenarios

With the continuous development of society,the living standard of people have been significantly improved.More and more families or enterprises have begun to own motor vehicles.The popularity of motor vehicles has provided us with a lot of convenience.However,with the increase in the number of motor vehicles,road traffic conditions have become more and more complicated.At the same time,incidents such as vehicle decking,red lights,accidental escape and other public safety,endangering road traffic safety,have begun to occur frequently,which brings great security risks to our travel.To solve the traffic safety problem,we need to effectively identify the motor vehicle.At present,the license plate information cannot be used as the unique identifier of the vehicle.We also need to identify the color,model and even the position of the vehicle in the monitoring probe.The rapid development of deep learning provides new solutions to these problems.This thesis uses the deep learning algorithm to solve the problems of vehicle identification and detection.The main research contents include the following two aspects:The design of vehicle type recognition based on capsule network is advanced and implemented in this thesis.The capsule network is a new algorithm proposed by the team of Professor Geoffrey Hinton to replace the application of convolutional neural networks in pattern recognition.The algorithm focuses on one feature with a set of neurons as a capsule.The adjacent two layers of capsules are connected by a dynamic routing algorithm.The underlying capsules form a prediction vector by affine transformation to predict the high-level capsules.The training sample contains five common models from different shooting angles and lighting conditions.In the process of training,in order to improve the convergence rate and recognition rate of the model,this thesis optimizes the dynamic routing algorithm of the capsule network.Comparing with the original model,the proposed model maintains a high recognition rate and shortens the training model time.The average recognition rate of the five models reaches 91.27%.This thesis is based on the Faster RCNN structure for vehicle type detection.In the process of vehicle type detection,the features extracted by the convolution layer have a great influence on the detection accuracy.This thesis implements a vehicle detection scheme based on the multi-feature fusion of Faster RCNN structure.The up-sampling method is used to merge the feature maps of different layers of the convolution sharing part,so that each layered feature map contains more context information.Using this feature fusion method,the effect of vehicle type detection has been significantly improved without substantially increasing the computational overhead.After experimental comparison and analysis,the accuracy of the proposed model is 3.27% higher than that of the original Faster RCNN model.