Search papers, labs, and topics across Lattice.
This paper introduces CylindTrack, a novel depth-aware cylindrical tracking-by-detection framework specifically designed for panoramic multi-object tracking (MOT). By implementing Depth-Temporal Trajectory Modeling (DTM) and Spherical Spatio-Temporal Consistency Learning (SSTC), the framework enhances depth observations and maintains trajectory continuity in the complex periodic domain of panoramic videos. The results demonstrate significant improvements in identity preservation and tracking accuracy, addressing the unique challenges posed by panoramic scenes with high object density and varying scales.
CylindTrack achieves superior identity preservation in panoramic multi-object tracking by effectively modeling depth consistency and leveraging the unique topology of equirectangular images.
Multi-Object Tracking (MOT) is a core capability for embodied perception, and panoramic cameras are attractive for embodied systems because their 360{\deg} field of view reduces blind spots and keeps surrounding targets observable for longer durations. However, panoramic MOT is not a straightforward extension of perspective MOT. In equirectangular panoramic videos, the horizontal image domain is periodic rather than Euclidean, which breaks planar motion assumptions and makes IoU-based association unreliable near the 0{\deg}/360{\deg} seam. Meanwhile, large-FoV scenes often contain more objects, stronger scale variation, and more frequent interactions, making online association particularly sensitive to unstable frame-wise depth cues. To address these issues, we propose CylindTrack, a depth-aware cylindrical tracking-by-detection framework for panoramic MOT. CylindTrack first introduces Depth-Temporal Trajectory Modeling (DTM), which promotes instance depth from an isolated frame-wise cue to a temporally filtered trajectory-level state. To improve the reliability of depth observations, we further develop Spherical Spatio-Temporal Consistency Learning (SSTC), which combines a Temporal Mixer and Spherical Geometry-aware Attention to enhance temporal coherence and panoramic geometric alignment in depth-aware representations. Finally, we design a Topology-Aware Cylindrical Motion Model (TCMM) that lifts horizontal motion into a continuous angular state space and performs seam-consistent motion prediction and association in the periodic panoramic domain. By jointly modeling trajectory-level depth consistency and panoramic topology, CylindTrack improves identity preservation and trajectory continuity in challenging panoramic scenes. The source code will be released at https://github.com/warriordby/CylindTrack.
