3D Single Object Tracking Based on Polar Coordinate Representation and Foreground Attention Mechanism
The current mainstream method for 3D single-object tracking involves tracking the object by matching features be-tween a template and a search area,making feature extraction and matching of the template and search region crucial.However,point cloud features extracted by current mainstream methods exhibit orientation sensitivity,which can adversely affect subsequent feature matching and detection.To alleviate these issues,this paper proposes a 3D single-object tracking network based on polar co-ordinate representation and foreground attention mechanism.The model helps the network learn rotation-invariant point cloud fea-tures through polar coordinate space transformation.Moreover,the scarcity of foreground points can lead to the tracking model fail-ing to adequately focus on the foreground object.For this reason,a foreground weight enhancement module is proposed that enables the 3D tracking model to better track foreground objects.Extensive experiments conduct on the KITTI and NuScenes datasets demon-strate that our method achieves improvements of 3.9%,7.6%in precision and 2.5%,8.0%in success rate over baseline methods for the car category.
3d single object trackingpolar coordinate feature representationforeground attentiondeep learning
万方数据
维普
