The Research And Implementation Of Vehicle Class And Model Recognition Under Multiple Angles

With the development of the continuous improvement of the people’s consumption level,the number of vehicles is increasing.The real-time vehicle monitoring management is difficult to complete by manpower alone,and intelligent traffic with artificial intelligence must be used.The classification of vehicle brand and models is an important part of vehicle monitoring management.People can analyze the video images under the surveillance cameras on the road to realize vehicle model classification,and then carry out statistics and data analysis for vehicles in the city effectively.In recent years,the field of computer vision has flourished.In many visual processing tasks,the recognition level of computers has even exceeded the normal level of the naked eye.However,in the actual traffic environment,the angles captured by the surveillance cameras are various,and the external conditions such as shooting time,weather,and illumination angle have a great influence on the picture clarity of the captured pictures,which makes the classification of the vehicle model more difficult.Also,it is very difficult and takes a lot of work for manual labeling.Which is more,The actual traffic environment varies widely,and the differences in different environments are large.In a new environment,the model with better classification effect may have a larger accuracy gap.How to maintain a high accuracy when the environment changes is relatively difficult.The purpose of this paper is to classify the brand and models information of the vehicles in the case that the surveillance camera has detected the vehicle and intercepted it into a picture,and solves the problem of cold start of vehicle brand recognition in the new environment,so that it can maintain high accuracy for vehicle model classification from different angles and environments.The main work of this paper contains:(1)Fine-tuning the internet source vehicle image using different convolutional neural networks,to verify the effectiveness of using the convolutional neural network for fine-tuning training to solve such problems of vehicle model classification under multiple angles.(2)Collect and produce a large-data-quantity surveillance camera-sourced vehicle image datasets marked by vehicle models to fill the gaps in the vehicle brand identification field where there are few vehicle image datasets for surveillance cameras.(3)Use the domain adaptation method in transfer learning.Transfer the vehicle model classification model of internet source vehicle image fine-tuning training to the surveillance camera source dataset,so that keep a high recognition accuracy in a cold start period without vehicle model data is obtained in the target environment.