Research On Vehicle Multi-feature Recognition Method Based On Deep Learning

With the continuous development of the economic level and the continuous development of the increase in the number of vehicles,it is urgent to build an intelligent transportation system.As an important input to the intelligent transportation system,road monitoring contains the multi-dimensional characteristics of vehicles.At present,only the license plate information of vehicles is widely used,and it is impossible to solve the problem of unlicensed vehicles and vehicle occlusion.Under the circumstances,vehicle identity recognition and vehicle multi-feature recognition have practical application value for the construction of intelligent transportation.In recent years,the development of deep learning has strengthened the ability of vehicle feature extraction in road monitoring.Vehicle multi-feature recognition based on deep learning has a new development opportunity.The multi-features of vehicles include vehicle color,type,vehicle logo and license plate number,etc.It is difficult to obtain the vehicle logo and license plate number characteristics of high-speed vehicles under road monitoring.This paper identifies the color and type characteristics of the vehicle.Because of the road monitoring With complex background,changeable illumination,vehicle occlusion and motion blur,etc.,using a network to identify vehicles with multiple features has low accuracy.This paper decomposes the multi-feature recognition task into vehicle detection task and multi-attribute recognition task,and conducts algorithm research on the two tasks.The main work of this paper is as follows:Based on the research of improved YOLO road monitoring vehicle detection algorithm,in order to avoid the influence of the initial anchor frame on the selection of the vehicle detection anchor frame in the YOLO network,and make the anchor frame more match the size of the vehicle,the K-Means adaptive anchor frame of the quantum genetic algorithm is proposed.The selection method uses quantum coding and quantum update methods to improve the convergence speed and accuracy of the anchor frame;the introduction of a cross-stage local feature pyramid module,through hierarchical feature fusion,improves the network’s receptive field,and reduces the amount of calculation while ensuring the detection accuracy.Good results have been verified by experiments on UA-DETRAC and extended data sets.Research on vehicle multi-attribute recognition algorithm based on Bi-Linear CNN.Aiming at the problem of fine-grained multi-label classification of vehicle multi-attribute recognition,the Bi-Linear CNN model is adopted.The specific work is as follows:(1)Study the impact of the dual-stream branch network model structure on the accuracy of vehicle multi-attribute recognition,use Res Net and VGGNet to construct three combined bilinear convolutional neural network structures to subdivide the vehicle multi-attribute classification.(2)Study the impact of tasks in vehicle multi-attribute recognition.By adjusting the weighting factors of vehicle type loss function,consider the relationship between vehicle type attributes and vehicle color attributes.(3)Aiming at the problem of low accuracy of vehicle multi-attribute recognition for rare categories,a category equalization factor is introduced into the loss function to strengthen the importance of rare categories in network learning.The vehicle multi-attribute recognition algorithm is tested on the custom data set intercepted by UADETRAC and the Ve Ri data set to verify the effectiveness of the Bi-Linear CNN model in vehicle multi-attribute recognition,and explore the relationship between vehicle type attributes and vehicle color attributes.It proves the superiority of class equalization factor in vehicle multi-feature recognition.