Research On Vehicle Instance Segmentation Algorithm Based On UAV Aerial Images

The rapid development of road traffic brings convenience to people’s work and life.At the same time,the increasingly complex road conditions also bring many hidden dangers for travel safety.Serious traffic congestion and frequent traffic accidents bring severe challenges to the traditional way of traffic supervision.It is of great significance to improve the traffic monitoring system to ensure road safety and relieve traffic pressure.With the advantages of high mobility and flexibility,UAV has been widely used in traffic monitoring system.UAV aerial images contain abundant traffic information,so it is of great significance to process the aerial images to extract effective information automatically.Vehicle instance segmentation based on UAV aerial image is an important part of traffic monitoring system,which plays an important role in traffic flow monitoring,highway inspection,traffic accident forensics and so on.Compared with the general case segmentation method,the aerial vehicle case segmentation method faces many challenges,such as lack of training data,arbitrary orientation of vehicle,large scale change of vehicle and complex image background.In this paper,the vehicle segmentation algorithm based on UAV aerial image is studied systematicallyFirstly,in view of the lack of data in the field of aerial vehicle instance segmentation,this paper constructs an aerial vehicle instance segmentation data set based on UAV,and summarizes a computer-aided annotation method,which can effectively reduce the workload in the process of data annotation.Secondly,to solve the problem of vehicles in arbitrary direction in UAV images,this paper propose an instance segmentation method based on active convolution kernel.Based on the Mask R-CNN,the active convolution kernel is used to replace the traditional 2D convolution kernel,and the original classification branch and regression branch are separated The rotation sensitive features extracted by active rotating filters are used for regression branch and mask branch,global average pooling is applied to rotation sensitive features to obtain rotation invariant features,which are applied to classification branches.Thirdly,aiming at the problems of large scale change of vehicles and complex background of aerial images,this paper propose an aerial vehicle instance segmentation method based on attention mechanism.In the backbone network,the channel attention mechanism is applied to make the high-level semantic features in the deep network and the detail features in the shallow network fuse adaptively according to their importance,so as to obtain more balanced multi-scale features.In the head network,the classification branch and the regression branch are separated,and the spatial attention mechanism is applied to the position sensitive regression branch and the mask branch to guide the head network to pay more attention to the target vehicle area and suppress the background noise.The proposed method is verified on the challenging UVSD dataset,and the experimental results are good and better than the same type of algorithm,which verifies the effectiveness of the proposed algorithm.