Vehicle Targeting Based On Visual Attention And Deep Learning

The development of science and technology has driven the rapid development of China's modernization process,and the intelligent transportation system has attracted more and more attention from institutions.As the main basic link of intelligent transportation,vehicle target detection has received more and more attention from researchers.Visual attention mechanism can simulate the human visual structure,focus on the prominent target area and ignore the background information,so as to be able to prioritize the target information in complex external environment.Deep learning can extract features on a layer by layer basis,the high level uses the underlying features as input,extracts higher level feature information,and receives researchers' attention in the field of image classification and target detection.This article mainly completed the following work:1)Image feature point matching is widely used for target detection.The image feature point can not fully reflect the image information in a single scale,this paper proposes a matching algorithm that combines multi-scale feature points.The feature points of the image at different scales are merged,and then the feature points are matched.Experiments show that the fusion of multi-scale feature points has higher matching accuracy than single-scale feature point matching algorithm.The visual attention mechanism algorithm can exclude some background information and reduce the interference of irrelevant information;The depth image reflects the information related to the position and spatial relationship and ensures the accuracy of the extracted target region;The depth image and saliency map obtained through visual attention mechanism are merged to get more accurate salient areas,then match the feature points for the salient areas.Experiments show that the algorithm has better matching accuracy and shorter matching time.2)Train vehicle classifiers with convolutional neural networks,apply back propagation algorithm to adjust network parameters;Use different network models CaffeNet,VGGNet,GooGleNet to train classifiers,and compared the classifiers trained with traditional machine learning algorithms SVM,Adaboost,and ANN.Experiments show that deep learning has a higher recognition rate than traditional machine learning algorithms.Then,uses Selective Search algorithm to extract candidate regions from the source image and submit it to the trained model for vehicle target detection.3)The complex background information has a great influence on the vehicle detection.Firstly,the visual attention mechanism can focus on a few salient areas;Then the background prior and foreground prior information of the image is used to extract the high-quality significance area,and the vehicle contour information is used at the same time.The significant areas are filtered to obtain more accurate salient areas;finally,the extracted salient areas are submitted to the classifier for vehicle target recognition.Experiments show that the image processed by the visual attention mechanism is greatly improved compared to the detection rate of the vehicle object before processing;and when the external environment changes in light intensity,rotation,etc.,it can maintain good stability.