Research On Vehicle Target Detection And Tracking Algorithm In Aerial Video

In recent years,the huge increase in the number of vehicles in China has led to increasingly complicated road conditions.It is important to improve the traffic monitoring system to ensure the road safety.And vehicle detection and tracking is an important part of traffic monitoring.Compared with the ground monitoring mechanism,the air monitoring mechanism covers a wider range and can obtain more road information.Moreover,the vehicle usually appears in any direction,which makes the traditional target detection method based on horizontal bounding box unable to effectively identify the vehicle information.Therefore,this paper mainly focused on vehicle detection and tracking in aerial photography based on rotating bounding box.With the development of deep learning,the object detection method based on convolutional neural network has become the mainstream of the object detection field.This paper studied several current mainstream object detection algorithms.Through comparing their detection effects on public remote sensing dataset DOTA,the possibility of applying Faster RCNN to aerial vehicle detection was analyzed.For the detection of aerial vehicles,this paper selected UCAS-AOD aerial data set for network training and testing,and used data augmentation to expand the training set.This paper firstly carried out the changes of the design of the rotating anchor frame,IOU calculation and the structure of ROI Pooling on Faster R-CNN to construct a detection network of vehicle rotating target.This network realized the detection of the rotating bounding boxes of aerial vehicles.Then,in order to improve the accuracy of vehicle target detection,the deep convolutional neural networks ResNet network and MobileNetV2 network were used to replace the original basic feature extraction network.Experimental results show that,compared with the original network model,the detection accuracies of the two replaced network models were both improved.And through experimental comparison and analysis,the network model based on ResNet with better detection effect was selected for further improvement.Finally,aiming at the problems of background misjudgment and vehicle missed detection,the attention model and multi-scale feature fusion were used to further improve the network model,making the vehicle detection results more accurate.For aerial vehicle tracking,this paper firstly combined the trained vehicle rotating target detection network with SORT multi-target tracking algorithm to realize the multi-target tracking of vehicles in aerial video.Then,aiming at the problem of ID switching after the same vehicle target turns,a data association rematch algorithm based on the combination of motion information and deep feature information is proposed.In the improved algorithm,the vehicle detection network was firstly used to detect and identify the video frames to obtain the rotating bounding boxes and deep feature vectors of the vehicle targets.And the Kalman filter was used to predict the rotating bounding boxes of the vehicle targets.Then,the Hungarian algorithm for IOU data association matching was used and the proposed data association rematch algorithm was used to match the vehicle detection results of the first pairing failure with the tracker a second time.Finally,the vehicle tracking results were obtained.Experimental result shows that the improved multi-target tracking algorithm can greatly solve the problem of ID switching after turning on the basis of vehicle tracking.