3D bounding box detection of moving objects for robot navigation in dynamic environments

Abstract

3D object detection is widely used in the field of robotics to avoid collisions with dynamic objects, such as humans. Based on the trajectories of the detected objects and the noise of the detection pipeline, robots can use motion planning algorithms to safely navigate in crowded environments. This thesis proposes a hybrid approach to 3D object detection using depth data fused with a mature 2D object detector. Both simulated and real-world experiments were performed as a demonstration of the solution. An extensive noise analysis of the developed detection pipeline was also carried out for future use in robot navigation under uncertainty.