| In recent years,with the rapid development of science and technology and the continuous improvement of peoples living standards,private car travel has gradually become the most popular way of travel.The Advanced Driving Assistance System(ADAS)plays an important role in driving.The detection and tracking of moving vehicles are very important for ADAS.However,due to the complexity of driving environment,occlusions between vehicles,illumination variation and so on,the detection and tracking of vehicles are still facing great challenges.Focusing on these problems,this paper mainly studies the following contents.Detection of small vehicles.The existing detection methods could not achieve a better detection of small vehicles in a long distance.In high-speed driving,the two vehicles would meet very quickly.Therefore,we hope that the detection algorithm can detect the small vehicles in a long distance,so that the moving vehicle can always keep a proper safe distance from the vehicles ahead.Based on this,this paper proposes a small vehicle detection algorithm based on multi-scale image group and the selection of region proposals.Firstly,the multi-scale image group is fused into the deep neural network to make full use of the spatial information of the image.Then,a new method of calculating the overlap between the ground truth and the proposals is designed to make the candidate proposals closer to the size of the vehicles to be detected.Finally,a loss function is designed to make the positive samples take up a larger proportion in the training process.In addition,in order to verify the effectiveness of the proposed algorithm,a dataset for small vehicle detection is constructed,and the experimental results prove the effectiveness of the proposed algorithm.Tracking of occlusion vehicles.Occlusion is a challenging problem in visual tracking.When the tracked vehicle is occluded by other vehicles,the tracking algorithm can easily miss the object and result in tracking failure.On the road,there are many occlusion cases.Therefore,it is very important for vehicle tracking to solve the occlusion problem.Based on this,this paper proposes a tracking algorithm based on filter selection.Firstly,in the stage of feature extraction,the filters which are more sensitive to the tracked vehicle are selected to represent the tracked vehicle.Then,the Discriminative Siamese Network(DSN)is designed to evaluate the tracking result.Finally,the tracked vehicle will be detected when tracking failure happens.In addition,we build a vehicle tracking dataset to verify the effectiveness of the proposed algorithm.Acceleration of vehicle detection and tracking.The popularity of deep learning makes vehicle detection and tracking achieve better results in accuracy.However,deep neural networks take a long time to extract features,which hardly meet the real-time requirements and practical application.We believe that to put the vehicle detection and tracking method based on deep learning into industrial use,we need to speed up the algorithm.To solve the speed problem of deep learning method,we reduce the parameters of neural network by adjusting the structure of the network.Experiments show that the neural network model can effectively improve the detection speed. |
