Research On Vehicle Detection Based On Feature Parallel Fusion

With the socio-economic development,the contradiction between limited road resources and the increasing number of vehicles has gradually emerged.Nowadays,with frequent traffic accidents and road congestion,autonomous driving technology has become a hot research direction.Among them,vehicle detection is an important part of environmental perception in autonomous driving,which is related to the safety and comfort of the vehicle during driving.Therefore,vehicle detection is taken as the research content of this paper.This paper first designs a shallow network and regression function to extract the region of interest,and then verifies the effectiveness of various artificial features and deep network features based on Triplet loss for vehicle recognition.The LBP features and SURF features are selected as artificial features,combined with deep network features,are the main features of this study.Aiming at the artificial features,a Gaussian inverse sampling method based on intraclass distance is proposed to model the contingency of the samples.For deep network features,the contingency modeling of the samples is performed by reconstructing the Triplet loss function method.Finally,artificial features and deep network features are combined in parallel in the form of complex features.Aiming at vehicle complex features,a hybrid discriminant analysis based on principal component analysis,linear discriminant analysis,and canonical correlation analysis is proposed for feature selection.This method selects the original complex features by considering the variance of the complex features,the ratio of the distance between the classes to the distance within the classes,and the correlation between the real and imaginary parts of the complex features.Aiming at the complex features after feature selection,a complex support vector machine algorithm is proposed to classify the complex features.The algorithm finds a hyperplane that can simultaneously divide the positive and negative samples of the real part and the positive and negative samples of the real part at the maximum interval by reconstructing the objective function of the support vector machine,and then optimizes the loss function by stochastic gradient descent to find this hyperplane.Finally,by combining the ideas of the random forest algorithm,a single complex support vector machine classifier is extended to a classifier group with multiple complex support vector machines to support the recognition of multi-class vehicle targets.The method first divides the sample set into multiple sub-sample sets randomly to train multiple classifiers,and then weights the output of multiple classifiers to obtain the final result of the classifier group.The optimal weighting coefficient is determined by the support vector machine algorithm.