Research On Vehicle Detection Algorithm In Aerial Image Based On Deep Learning

With the acceleration of urbanization,smart cities and smart transportation have become an inevitable trend in the future.Vehicle detection and classification play an important role in these two smart conceptions.In recent years,with the rapid development of drone technology,vehicle detection technology on drone-based aerial image has gradually been widely used.Aerial images are token by drones at high altitude.However,the vehicles in aerial image have many difficult factors,such as low resolution,dense and complex scenes.The detection performance through traditional feature extraction and detection methods is unsatisfactory.Due to these problems,this thesis uses the deep learning to do the following researches.Firstly,the current methods which use bounding box can’t get excellent performance in some scenes,i.e.,limited resolution and dense distribution,so a novelty method based on heat map is proposed.In this thesis,a two-dimensional elliptic function is adopted.It uses the angle,center position,length,and width of the vehicle to generate a heat map and treat it as a ground truth.Because of using the heat maps,a series of complex algorithms for general object detection based on deep learning can be discarded,such as anchor generation algorithms and maximum suppression algorithms.The experiment results show that the performance of using heat map is better than these bounding-box methods.And this thesis use two different methods from the algorithm level to reduce the performance impact caused by the object imbalance problem.Secondly,in order to enable the model to can be used in limited memory’s devices,such as drones,a fully convolutional lightweight pyramid network is proposed.Down-sampling and gradually deepening channel are used to get sufficient semantic information.Gradually up-sampling and skip-connection on feature maps of the down-sampling stage are used to combine semantic information and spatial information to improving the detection performance.The proposed models in this thesis do not use any pre-training network,fully connected layer and classifier.Experimental results show that the proposed model in this thesis is not only lightweight,but also get excellent detection performance.Finally,according to different application scenarios,such as detection and classification,this thesis can use single-channel and multi-channel heat maps,respectively.In the scenario of vehicle classification,a Refine Module is introduced to fuse multiple scale information,experimental results show that the refining module improves the detection and classification performance of vehicles in aerial images.