Research On Road Reconstruction Technology Based On UAV Aerial Images

With the rapid development of artificial intelligence,driverless technology gradually comes into the public consciousness and becomes a hot research subject.HD maps,as a significant module of driverless cars,can provide much more abundant and precise information than traditional maps,which means the traditiornal maps is not adaptive for driverless car anymore.Therefore,efficient collection and reconstruction methods for HD maps are extremely urgent.The existed collection method for HD maps relies on vehicular cameras.However,images from vehicular cameras suffer from narrow vision field and impact of perspective,and the route of car is uncontrollable because of complex traffic conditions.Thus,we propose the collection method of HD maps based on unmanned aerial vehicle and carry out research into 2D and 3D reconstruction project for urban road scene,taking account of low texture and high similarity in road scene.For 2D reconstruction of road scene,we propose the image stitch algorithm based on sparse optical flow,which could acquire the panoramic image from high solution aerial images in large-scale scenery.We track features in image sequence with Lucas-Kanade optical flow in order to achieve image alignment.Then we stitch images selected as key frames in the global map-based stitching framework.Before fusing the overlap area,we would rectify the transform between the global map and key frames once more to reduce the error accumulated over time.For 3D reconstruction of road scene,we propose the planar-like scene reconstruction algorithm based on semi-direct method,simultaneously reconstructing the sparse 3D structure of road and the trace of camera in near real-time.We employ semi-direct method to match 3D points with features extracted from images.Under the planar hypothesis,we restrain the points on the plane,thus reduce variables in bundle adjustment to a small set,which makes the algorithm much more efficient.At last,we apply the trace of camera into image stitch algorithm,and successfully get the image of the whole road.Compared with satellite image,our image has much higher ground resolutions achieving 1cm/pixel.With these high-resolution images,we can build HD map quickly.