A Moving Object Detection Method for Mobile Robots Using Feature Point Sticky Tree
20 Pages Posted: 27 Dec 2021
Abstract
Moving object detection is a common task for mobile robots when working in dynamic environments. Due to the nonstatic nature of the mobile robots themselves, we need to conduct the motion detection work under the moving camera assumption. Feature-point-based object detection is an important research direction to solve this problem. The distances and motion conditions of the feature points in both 2D image space and 3D physical space are critical criteria to cluster the points to objects. However, previous work never fully considered all the criteria and explored their complex relationship with object segmentation outputs. As a result, accurate segmentation can not be achieved when the differentiation condition only presents in part of the criteria space. Therefore, we proposed a novel feature point stickiness method to include all the above considerations. A fuzzy neural network is designed to assess the stickiness between any feature points, e.g., the possibility of being on the same object. The network is trained with small data set to capture the complex nonlinear relationship between the point distance and motions and the stickiness. In addition, a sticky tree growth algorithm is used to connect the feature points according to their stickiness distributions. The formation of the trees autonomously completes the object segmentation task. The method is testified in various experimental situations, including some challenging scenarios. It improved the object recall rate and precision rate to a large extent compared to previous works and showed high potential in providing precise trajectories of moving objects in the robot’s workspace.
Keywords: moving object detection, feature points, sticky tree, moving camera
Suggested Citation: Suggested Citation