Vehicle Detection And Vehicle Type Identification Algorithm Based On Convolutional Neural Network

As a key technology of advanced driving assistance system,vehicle detection and vehicle-type identification plays a vital role in reducing traffic accidents,protecting the lives and property of personnel,and realizing intelligent transportation.It is also an important part of intelligent transportation system.However,affected by the unfavorable factors in complex traffic scenarios,such as multi-illumination,multi-angle and multi-scale,the performance and robustness of vehicle detection and vehicle-type identification are poor.This paper aims at the vehicle detection and recognition scene of monocular front-view in the vehicle-aided driving system,focusing on the issue about vehicle detection,localization rapidly and robustly,and vehicle-type identification precisely in complex scenes,such as multi-illumination,multi-angle and multi-scale.The main research contents are as follows:1.Aiming at the problem of front-view target localization,a multi-scale front-view imaging localization model(MSFI-LM)based on deep convolution network is studied.First,based on the inverse perspective mapping model,the external parameters of vehicle camera is calibrated by the experiment.Then,the accurate airspace front-view imaging model is established.Based on this model,the localization formula of front-view vehicle is derived and mapped to multi-scale convolution layers,and then MSFI-LM is established.The measured data shows that,the positioning model has high positioning accuracy,which lays the foundation for the subsequent integration of vehicle detection and localization.2.Aiming at the problem of front-view variable-scale vehicle detection and localization,a fast algorithm for front-view variable-scale vehicle localization before detection is implemented.First,Combining the vehicle characteristics of monocular front-view scene,the convolutional neural network is fused with the prior information of airspace in multi-scale.At the same time,the structure of convolutional neural network is improved,then an embedded airspace convolution with multi-scale joint detection method is proposed.The problem of low computational efficiency caused by full-map convolution is avoided.Then,aiming at the problem of executing the locating calculation repeatedly in the traditional system of detection before localization,according to the multi-scale front-view imaging location model,the distance-image at each scale is calculated in advance.At the same time,the embedded airspace convolution multi-scale detection network and distance-image are combined deeply,a fast algorithm for front-view variable-scale vehicle localization before detection is implemented.In the video vehicle localization,So that only one single locating calculation process needs to be executed.Verified by measured data,the algorithm improves the efficiency of detection and location calculations effectively while maintaining high detection performance.3.Aiming at the problem of low recognition rate of front-view variable-scale vehicle in low-resolution images,a typical vehicle-type recognition algorithm based on transfer learning is implemented.the structured and parametric vehicle-type database is constructed.Then the transfer learning based on the embedded airspace convolution multi-scale joint detection architecture is implemented,In order to obtain the high-precision vehicle-type identifier,the dynamic learning strategy based on difficult sample recombination is used to train the network.This algorithm improves the recognition rate of front-view variable-scale vehicle in the low-resolution images effectively.Experiments show that the recognition performance of the algorithm is better than other algorithms.