Research On Vehicle Tracking And Segmentation From The Perspective Of UAV

Vehicle tracking and segmentation,which aims to track and segment the vehicles in image sequences,is a challenging problem in the traffic detection area.This technique is widely used in the visual tasks of UAV and provides a reliable data source for many other areas,such as counting the number of vehicles,behavior analysis,abnormal event detection in highways.Most of the existing researches try to extract more image features to improve the tracking and segmentation accuracy.However,due to the flying altitude of UAV,the images taken by UAV are consisted of small number of pixels and a little features.This makes the performance of the existing methods not ideal.At the same time,for the segmentation problem,the low resolution often leads to the adhesion between targets and the wrong segmentation of vehicle edges.To address the issue of unsteady vehicle tracking from an aerial perspective.This study proposes a multi-vehicle tracking model from the perspective of an unmanned aerial vehicle(UAV).To model the kinematic and behavioral properties of the target vehicle,the model employs a long-term and short-term memory network.The target’s historical motion trajectory is then used to estimate the next frame coordinates,and the trajectory of each target is anticipated and tracked in conjunction with the limitations and interactions of surrounding vehicles on the target vehicles.From the standpoint of a UAV,this strategy can significantly increase the accuracy and detection speed of multi-target tracking.Furthermore,we gathered a multi-vehicle tracking data set from the perspective of a UAV and recorded it in MOT16 format.The data set is used to train and test the proposed method and comparison algorithms.In comparison to the most advanced algorithms created in recent years,our suggested algorithm efficiently minimizes the frequency of missing targets,particularly in the event of limited visual feature discrimination,and has effectively enhanced the algorithm’s operating efficiency.To approach the previously mentioned segmentation problem,we propose an effective vehicle segmentation algorithm.The algorithm first quantifies the colour space based on the colour information of the pixels,then identifies the colour centres of the image and replaces the colour of each pixel with the colour centre.First,the photographs to be input are preprocessed to reduce the scatter in the image.Then,in the CIELab colour space,it is quantified by a clustering algorithm and irregular pixel clusters are obtained.Next,for the contrast prior,the statistical information of the colours of the input image is used to determine the significance values of the image pixel clusters.For the centre prior map,the centre of the valid target is first estimated using the target coarse localisation method.The saliency score of each pixel cluster is then calculated based on the distance of the cluster from the centre,resulting in a centre prior map.The contrast a priori map is then combined with the centre a priori map to obtain the initial saliency map.A post-processing step is then performed to obtain the final saliency map.In this paper,a set of vehicle segmentation data from the perspective of UAV aerial photography is collected to test the effectiveness of the proposed method.Also,to further demonstrate the generality and reasonableness of the proposed method in this study,this paper conducts comparative tests on a public dataset as well.The extensive experimental results show that the algorithm proposed in this thesis performs well on the vehicle segmentation problem.Finally,this paper combines the vehicle tracking and segmentation methods to propose a joint multi-vehicle tracking and segmentation simulation system.The system combines the two methods mentioned in the previous paper to segment vehicles while maintaining the motion trajectory of the target vehicles.The system can provide richer and more effective information for the field of intelligent transportation and is useful for traffic event detection such as vehicle collisions.