EE Seminar :NextStop: An improved tracker for panoptic LiDAR segmentation data
סמינר זה יחשב כסמינר שמיעה לתלמידי תואר שני ושלישי
Electrical Engineering Systems Seminar
Speaker: Nirit Alkalay
M.Sc. student under the supervision of Prof. Ben-Zion Bobrovsky
Sunday, 28th July 2024, at 15:00
Room 011, Kitot Building, Faculty of Engineering
NextStop: An improved tracker for panoptic LiDAR segmentation data
Abstract
4D panoptic LiDAR segmentation is essential for scene understanding in autonomous driving and robotics, combining semantic and instance segmentation of LiDAR point clouds with temporal consistency. Current approaches, like 4D-PLS and 4D-STOP, use a tracking-by-detection methodology, employing deep learning networks to perform semantic and instance segmentation on each frame. To maintain temporal consistency, large-size instances detected in the current frame are compared and associated with instances within a temporal window that includes the current and preceding frames.
The above mentioned association method rely on short-term instance detection within the temporal window, lack motion estimation capabilities, and exclude small-sized from matching, leading to frequent identity switches and reduced tracking performance.
To address these limitations, we introduce the NextStop1 tracker. Our tracker combines Kalman filter-based motion estimation, data association, and lifespan management modules, along with a tracklet state concept for prioritization. Additionally, we leverage accumulated tracking data to correct temporal inconsistencies in semantic segmentation results.
We evaluated our tracking method using the LiDAR Segmentation and Tracking Quality (LSTQ) metric, proposed by Aygun et al., on the validation set of SemanticKITTI.
NextStop showed improvements in this metric for various classes such as Other-vehicles, People, and Cars. The advantages of our tracking method lie primarily in tracking small size objects, including small-sized classes like People and Bicyclist, as well as objects from other classes that are at a distance and therefore considered small-sized. These improvements are reflected in fewer ID switches, earlier tracking initiation, and more reliable tracking in complex environments.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר
