Vehicle Detection Based On Drone Images With The Improved Faster R-CNN

With the improvement of economic strength,urban congestion and frequent traffic accidents are becoming more serious because of the higher vehicle penetration rate.The technology of vehicle detection can provide strong information support for the handling of many traffic problems such as road traffic control,highway management and emergency management,which has received increasing attention from researchers.The detection of vehicles in the drone images can be used to indicate the traffic conditions of the road and ensure the requirements of traffic control,and can also provide feedback on real-time road conditions to further protect pedestrians and drivers’ traffic safety.In order to solve the actual vehicle detection task,this paper studies the vehicle detection method based on drone image.The main research contents include:(1)A data set for the detection of unmanned aerial vehicle images has been constructed.This dataset contains all the drone images in the city and the vehicle tags in all images.The data set contains 800 images and 3320 vehicles are labeled.(2)In order to solve the false detection and missed detection problems in vehicle detection by the Faster R-CNN model,this paper improves the model by the following aspects: Firstly,select the better network residual network(ResNet101)to replace the VGG16.The model is used to improve the depth of feature extraction of convolutional neural networks;then the Feature Pyramids Network(FPN)are constructed using different hierarchical feature maps to further improve the breadth of network feature extraction,effectively solving the problem of detecting smaller targets of vehicles;The recommended regional scales in the network are optimized.Finally,based on the UAV image vehicle dataset,the improved Faster R-CNN network model was used to test the vehicle target detection.(3)In order to eliminate the influence of haze weather on drone image detection tasks,we compare different types of dehazing algorithms and analyzes the detection effect after dehazing.And compare the detection effect after using different algorithms to remove haze.Then the improved Faster R-CNN model is used to detect the vehicles in the image.The experimental results show that the detection effect is much better than the direct use of smog images after using different dehazing algorithms.The dark channel prior algorithm has the best detection effect after dehazing.