Accurate and effective monitoring of many objects in 3D space is critical in many
applications, including autonomous driving, robotics, and surveillance. The
increased availability of LiDAR sensors has enabled the capture of high-resolution
3D point clouds, which may then be used for object detection and tracking.
However, the complexity of 3D point clouds and the variety of object appearances
present substantial obstacles for multi-object tracking. In this thesis, we offer a
unique strategy for 3D multi-object tracking on LiDAR point clouds that combines
the capabilities of deep learning and classic computer vision methods.
As tracking by detection(TBD), we employ the PointVoxel-RCNN (PV-RCNN)
model For object detection, a state-of-the-art 3D object detector that extracts robust
features from point clouds. The PV-RCNN model is a two-stage detector that first
generates proposals using a region proposal network (RPN) and then refines the
proposals using a second-stage network. To correlate detections across frames, we
use a greedy matching method with the Mahalanobis distance. The Mahalanobis
distance is a metric that calculates the distance between two points in a space with
multiple dimensions while taking into account their covariance. To calculate the
level of similarity between detections in subsequent frames, we use the
Mahalanobis distance. Our method allows for the uncertainty of both detection and
tracking, ensuring stability in the presence of occlusions, noise, and variable point
cloud densities. We also employ a Kalman filter to predict object motion and
improve tracking accuracy. The Kalman filter is a model of mathematics that uses
noisy data to determine a system's state. It is often utilized in tracking applications.
The value of our technology is demonstrated by the experimental results on the
KITTI tracking benchmark, which show competitive accuracy and precision. The
KITTI benchmark is a well-known testing environment for 3D object detection and
tracking, consisting of 21 LiDAR point clouds and camera pictures. Our system has
a tracking accuracy of 76.51% and a precision of 87.42%.