Leveraging Stereo-Camera Data for Real-Time Dynamic Obstacle Detection and Tracking

Abstract

Dynamic obstacle avoidance is one crucial component for compliant navigation in crowded environments. In this paper we present a system for accurate and reliable detection and tracking of dynamic objects using noisy point cloud data generated by stereo cameras. Our solution is realtime capable and specifically designed for the deployment on computationally-constrained unmanned ground vehicles. The proposed approach identifies individual objects in the robot's surroundings and classifies them as either static or dynamic. The dynamic objects are labeled as either a person or a generic dynamic object. We then estimate their velocities to generate a 2D occupancy grid that is suitable for performing obstacle avoidance. We evaluate the system in indoor and outdoor scenarios and achieve real-time performance on a consumergrade computer. On our test-dataset, we reach a MOTP of 0.07 ± 0.07m, and a MOTA of 85.3% for the detection and tracking of dynamic objects. We reach a precision of 96.9% for the detection of static objects.

Autonomous Flights in Dynamic Environments with Onboard Vision

“…This method divides the point cloud into individual objects without dynamic classification, which does not represent the dynamic environment accurately. Our approach builds upon and extends a dynamic obstacle detection and tracking algorithm [1], which tracks clusters and classifies them as dynamic or static. Varying from them, we conduct tracking with occlusion-awareness and replace their human detector with our proposed re-free strategy to address the issue of erroneous occupancy caused by temporary standstills.…”

“…In this section, we describe a dynamic perception module built upon [1]. Given a point cloud generated from depth image at time τ ∈ R, we filter it to reduce sensor noise and lower computational overhead, and then cluster the filtered point cloud into individual objects per frame using DBSCAN [31], resulting in a set of m clusters…”

“…We denote a list of cluster sets over k frames separated by ∆t as C t,k = {C t−k•∆t ,..., C t−∆t , C t }. Given C t,k , we use an occlusion-aware tracking algorithm extended from [1] to obtain their tracking history…”

“…In this paper, we bridge this research gap by introducing a complete system composed of an accurate dynamic environment perception module and an accompanied trajectory generation method. For dynamic environment perception, we classify the obstacles into static ones and dynamic ones by clustering, tracking, and voting the point cloud data from a depth sensor, inspired by [1]. To address the issue of occlusion, we propose an occlusion-aware Kalman filter to estimate the motion states of dynamic objects, including position, velocity, size, and associated uncertainty, making the classification more accurate and robust.…”

“…1) We develop an occlusion-aware dynamic environment perception method based on [1], which efficiently yet accurately classifies the obstacles into static ones and dynamic ones. Moreover, we represent both classes of obstacles as two different data structures to handle generic complex environments.…”