Remote Sensing Vehicle Detection Based On Artificial Neural Network

In recent years,with the rapid development of the economy,the number of domestic motor vehicles has gradually increased and traffic congestion has become increasingly serious.A wide range of vehicle detection and identification can timely obtain traffic information,which is very helpful for real-time monitoring of road traffic,traffic management,and dispatch.The theoretical basis of artificial neural network has developed rapidly in recent years.Most of the problems in the field of image target detection have been solved using artificial neural network methods and have indeed achieved wide success.Target detection tasks based on artificial neural networks,especially based on convolution features,are more frequent.Therefore,the application of artificial neural network to remote sensing image vehicle detection has important research significance in improving traffic safety and accelerating modern intelligent transportation systems.Vehicle detection methods are mainly divided into manual design feature methods and learnable feature methods.The method of artificially designing features has made great progress in vehicle detection.Because its classifiers mostly use BP neural networks,the number of network layers is shallow and the parameters are few,which has the advantages of short time consumption and rapid detection.However,in the environment,light,shielding and other external factors,the detection effect is poor,missed detection and false detection more.The method of learning features is generally based on the convolutional neural network model.Because the convolutional neural network has a high number of layers and a large number of parameters,it has the disadvantage of long detection time,but its powerful ability of autonomous learning improves the detection accuracy.It has become one of the hottest technologies.This paper studies vehicle detection methods based on artificial design features and vehicle detection methods based on learnable features.The specific work is as follows:(1)Establish a vehicle database.Collect high-resolution satellite remote sensing images from multiple roads,blocks,parking lots,etc.,calibrate vehicle objects in the image as positive samples of the database,and randomly calibrate non-vehicle targets in the image as negative samples of the database.All positive and negative samples with different calibration scales were saved to a 48 X 48 pixel size image using the bilinear difference method,and the angle of each vehicle’s positive sample was randomly rotated before saving.(2)Two image feature extraction algorithms are proposed.The three image feature extraction algorithms of HOG,LBP and RCD are studied in detail.According to the existing problems of RCD algorithm and the idea of LBP algorithm,two image feature extraction algorithms,LFP and RCD(G),are proposed.(3)Remote sensing image vehicle detection based on artificial design features.The HOG,LBP,RCD and the proposed LFP and RCD(G)algorithms are used to extract the characteristics of the vehicle database and sent to the BP classifier for training and testing.The performance of the proposed new algorithm is verified based on the experimental results.(4)Remote sensing image vehicle detection based on deep convolution features.The MyVGGNet convolutional neural network adapted to this database sample was designed to train and test the vehicle database,and vehicle detectors were used for vehicle detection in remote sensing images.