Abstract
Detection and tracking of dynamic objects is a key feature for autonomous behavior in a continuously changing environment. With the increasing popularity and capability of micro aerial vehicles (MAVs) efficient algorithms have to be utilized to enable multi object tracking on limited hardware and data provided by lightweight sensors. We present a novel segmentation approach based on a combination of median filters and an efficient pipeline for detection and tracking of small objects within sparse point clouds generated by a Velodyne VLP-16 sensor. We achieve real-time performance on a single core of our MAV hardware by exploiting the inherent structure of the data. Our approach is evaluated on simulated and real scans of in- and outdoor environments, obtaining results comparable to the state of the art. Additionally, we provide an application for filtering the dynamic and mapping the static part of the data, generating further insights into the performance of the pipeline on unlabeled data.
Introduction
Autonomous robots, particularly Micro Aerial Vehicles (MAVs), must perceive and predict the movement of dynamic objects to navigate safely. However, MAVs face a significant challenge: they have limited lifting capacity, requiring lightweight, low-resolution sensors, and limited onboard computational power.
Standard 3D LiDAR sensors used on MAVs, such as the Velodyne VLP-16, produce "sparse" point clouds consisting of only 16 scan rings. At a distance, a human might be represented by only a handful of points, making traditional shape-based recognition extremely difficult. Furthermore, many state-of-the-art tracking algorithms require powerful GPUs that a small drone cannot carry. This research presents a pipeline that is both robust to sparse data and efficient enough to run on a standard mobile CPU.
Methodology
The proposed pipeline is designed to exploit the inherent grid-like structure of organized LiDAR scans. By treating the data as a 2D grid (16 rings x n horizontal points) rather than an unorganized 3D point cloud, the algorithms become significantly faster.
Segmentation via Dual Median Filters
Instead of using a complex neural network, the system uses two median filters with different kernel sizes applied to the distance readings of the scan rings. A "noise filter" (small kernel) removes tiny outliers, while a "background filter" (larger kernel) identifies the general environment. If the two filters return significantly different results for a point, that point is classified as a foreground object. This effectively isolates objects based on their width, which is a more reliable feature than height in sparse data.
Detection and Tracking
Segmented points are grouped using a modified Region Growing Clustering (RGC) approach that handles sparsity by checking Euclidean distances between neighbors, preventing the accidental merging of distinct objects in close proximity.
The system maintains "object hypotheses" using Kalman filters with a constant velocity model. To match new detections to existing tracks, the Hungarian method is employed with a cost matrix based on the Bhattacharyya distance. Notably, the tracker provides feedback to the detector: if a track already exists, the detector "relaxes" its constraints to avoid losing the track due to momentary sensor occlusion or sparse returns.
Experimental Evaluation
The approach was evaluated on three distinct types of data: the InLiDa (Indoor LiDAR Dataset), simulated environments created in Gazebo with up to 50 dynamic persons, and real-world MAV data recorded using a DJI Matrice 600 in a courtyard in Bonn, Germany.
Key Results
- Real-Time Performance: The entire pipeline runs in under 100ms per scan on a single core of an Intel Core i7-6770HQ, making it viable for real-time flight.
- Accuracy: On the InLiDa dataset, the method achieved a Multi-Object Tracking Accuracy (MOTA) of 0.562, significantly outperforming the baseline approach (-0.213) which struggled with human/robot distinctions.
- Static vs. Dynamic Classification: Every track starts as "static" and is only promoted to "dynamic" once its bounding box moves enough to no longer intersect with its starting position. This helps filter out "ghost" tracks caused by sensor noise on stationary objects.
Conclusion
This work demonstrates that by prioritizing width-based features and exploiting the structured nature of LiDAR data, it is possible to achieve high-performance tracking on resource-constrained MAVs. While the system currently struggles with erratic motion and complex occlusions, it provides a robust foundation for real-time autonomous perception. Future work will focus on incorporating explicit occlusion reasoning to further stabilize tracks in dense environments.
Summary updated