A social robot has to keep track of the people around it over time to plan its motion — a task handled by 3D multi-object tracking. For the LiDAR modality, trackers usually sit on top of a detector and link its detections across frames with a separate motion model or re-identification network. We instead show how a 3D LiDAR detector itself can be turned into a tracker by adding an instance-level contrastive objective on its intermediate voxel features. Each embedding is read from the voxels inside a detected box, and boxes are matched across frames by cosine similarity and the Hungarian algorithm, without a motion model. The detector backbone is fine-tuned together with this objective under a gradient rule that keeps detection accuracy intact while learning identity-aware features. Because the method uses only the voxel features that LiDAR detectors already produce, it is backbone-agnostic.
Second on MOTA and OSPA. LiDAR-only — no camera fusion, no motion model, no separate re-identification network.
Output of the released model on a JRDB validation sequence, seen from above with the robot at the centre. Box colour is the track ID. There is no motion model: every frame, detections are matched to the tracks in memory only by the cosine similarity of their embeddings. When a person is occluded and comes back, the same ID is restored from that memory — the ring marks such a re-identification.
Timeline: one row per annotated person, coloured by the track ID the tracker assigned to them in each frame (grey: not detected). A row in a single colour is a person who kept one identity; a change of colour is an ID switch. Click a person in the scene or a row to follow one track; dashed boxes are the ground truth; rings every 4 m.
Training and inference code, configuration files and evaluation scripts are available at github.com/HerMak56/CLIFT; pre-trained checkpoints are linked from the repository README.