We present a decentralized multi-view tracking system for scalable camera networks. Single-view detections are projected to bird's-eye-view (BEV) coordinates and fused by independent multi-view trackers, each managing a dynamic spatial zone. To support scalability across spatially partitioned camera networks, we introduce a decentralized layer for trajectory unification, utilizing a locking protocol and periodic track-merging operations termed global collate. This process is supported by a delayed-output buffering mechanism that retroactively applies history corrections. The configurable delay window absorbs asynchronous coordination latency and helps stabilize cross-zone identity consistency. We detail the design of dynamic zone polygons, consensus-based coordination, and the validation framework, which together maintain global ID consistency with minimal accuracy loss compared to a single-instance baseline in the evaluated configurations.
Figures & tables
Fig. 1: High-level architecture. Single-view trackers (outer arcs) run per camera and report smoothed leg points in BEV coordinates. One or more multi-view trackers operate on BEV detections within their zone polygons. A decentralized layer maintains overlapping BEV zones and coordinates global collate operations to maintain consistent track IDs across zones.
Fig. 2: Decentralized multi-view tracker main loop. BEV detections are processed by the local BEV tracker. Tracks whose hit count lies in the global-collate window are collected, and the instance attempts to acquire the shared global lock for all overlapping zones. If the lock is obtained, it runs global collate and then collate validation; all accepted and reverted merges are written back into the match-history database before the iteration completes.
Fig. 3: Dynamic zone polygon expansion. The multi-view zone starts as a small polygon around the first few detections (left). As new detections arrive outside the current polygon, the polygon expands to include them (middle). After sufficient coverage, the polygon stabilizes and represents the final spatial footprint of the zone in bird’s-eye coordinates (right).
Fig. 4: Zone polygon management in the decentralized tracker. For each incoming BEV detection (x,y) , the instance checks whether the point lies inside its current zone polygon. If not, it expands its polygon and updates the decentralized zone_id → polygon map. All instances then check for polygon intersections; when overlaps are detected, each instance updates its local overlap map so that neighboring zones are known for later global collate.
Fig. 5: Impact of global collate on cross-zone tracking. (a) Without global collate, a person moving across overlapping zones generates fragmented local tracks with distinct IDs. (b) With global collate, local tracks are matched and merged, yielding a single consistent global ID for the entire trajectory.
Fig. 6: Global collate decision flow. Candidate matches must first pass a minimum hit-ratio check. If the incoming ID is unique locally, the merge is accepted immediately. If the ID already exists (conflict), the system forces the swap—removing the old ID—only if the match cost satisfies a strict force threshold.
Fig. 7: Collate validation logic. For a set of input tracks, the instance sends a global collate request and examines the returned matches. For each matched pair it accumulates hit counts and cost sums, builds an old_tid → new_tid map, and then compares the validated cost of the old track vs. the candidate new track. Depending on this comparison, the system either reverts the global collate match or swaps the local track ID and updates the matches database.
Dataset
Configuration
HOTA ( ↑ )
IDF1 ( ↑ )
IDSW ( ↓ )
Dataset A
1 Inst. (All)
0.911
0.962
4
2 Inst. (Split)
0.897
0.953
4
3 Inst. (Split)
0.885
0.942
8
Dataset B
1 Inst. (All)
0.810
0.906
24
2 Inst. (Split)
0.809
0.908
26
TABLE I: Accuracy Comparison at 5.0 s Delay. Note that the 3-instance partition substantially reduces local multi-view redundancy (to only 3–4 cameras per instance), resulting primarily in localized intra-zone tracking dropouts rather than decentralized cross-zone matching failures (see also Table VI ).
Dataset
Configuration
HOTA ( ↑ )
IDF1 ( ↑ )
IDSW ( ↓ )
Dataset A (2 Inst.)
No Global Collate
0.635
0.706
48
w/o Response Delay
0.869
0.937
13
Full System
0.897
0.953
4
Dataset B (2 Inst.)
No Global Collate
0.735
0.855
64
w/o Response Delay
0.779
0.891
43
Full System
0.809
0.908
26
TABLE II: Ablation Study on Decentralized Components (2 Instances)
Dataset
Delay (s)
HOTA ( ↑ )
IDF1 ( ↑ )
IDSW ( ↓ )
Dataset A
0.0
0.869
0.937
13
1.0
0.887
0.947
10
2.0
0.896
0.951
9
3.0
0.897
0.953
4
5.0
0.897
0.953
4
Dataset B
0.0
0.779
0.891
43
TABLE III: Delay Sensitivity for Two-Instance Decentralized Runs
Dataset
Proposals
Accepted
Skipped
Forced
Collision Rej.
Validation No-op
Reversions
Failed
Dataset A (2 Inst.)
81
36 (44%)
24 (30%)
0
1 (1%)
20 (25%)
0
0
Dataset A (3 Inst.)
163
74 (45%)
68 (42%)
6 (4%)
12 (7%)
3 (2%)
5
0
Dataset B (2 Inst.)
88
34 (39%)
44 (50%)
5 (6%)
5 (6%)
0
2
0
TABLE IV: Global Collate Proposal Outcomes at 5.0 s Delay
Dataset & Config
GT Tracks
Crossing GT
Total Cross.
Mean Overlap Fr.
Dataset A (2 Inst.)
62
28
137
∼ 844
Dataset A (3 Inst.)
62
34
165
∼ 1155
Dataset B (2 Inst.)
45
25
59
∼ 547
TABLE V: Ground-Truth Boundary Traffic and Overlap Profile
Fig. 8: Total processing latency vs. number of zones ( N=40 cameras). Data is generated with 100% overlap (worst-case), yet latency improves as zones are added.
Fig. 9: Latency breakdown (100 tracks, 40 cameras). As zones increase, core tracking time (blue) decreases faster than decentralized coordination (orange) increases.
Multi-camera tracking with overlapping fields of view typically relies on centralized fusion, which creates computational bottlenecks that prevent deployment at scale. We present MV3DT, a fully distributed framework for real-time multi-view 3D tracking that achieves accurate identity propagation and occlusion recovery through peer-to-peer coordination, eliminating the need for central aggregation. Each camera node executes a lightweight modular pipeline comprising monocular 3D perception, distributed multi-view association, and collaborative fusion via lightweight messaging. MV3DT achieves 96.5% IDF1, 93.1% MOTA, and 94.6% MOTP on WILDTRACK, competitive with state-of-the-art centralized methods, and unprecedented 41.7% IDF1 and 50.9% MOTA on SCOUT while demonstrating superior scalability: sustaining 30 FPS on 100 cameras with <10ms inter-camera latency and only 2.2% communication overhead. MV3DT operates in a zero-shot regime given camera calibrations, requiring no scene-specific learning and making it directly deployable in new environments. These results establish MV3DT as a practical solution for real-time multi-view tracking in large-scale overlapping camera networks.
Byron Hernandez, Fangyu Li, Aotian Wu +3
University of Florida, Gainesville, FL, USA · NVIDIA Corporation, Santa Clara, CA, USA
Precise multi-camera tracking traditionally relies on rigorous 3D site calibration, yet this requirement is often operationally impossible in large-scale deployments. Privacy regulations frequently prohibit recording video for offline calibration; limited bandwidth precludes synchronizing high-resolution streams from hundreds of cameras; and covering immense physical sites with calibration targets is logistically infeasible. We present a multi-camera homography calibration system designed to overcome these barriers through "on-the-fly" geometric refinement. Starting from coarse manual homographies, we introduce a centroid-based projection optimization (PO) that continuously aligns the ground-plane geometry using live detection streams. Because PO operates asynchronously on already-transmitted, lightweight metadata, it adds zero computational latency to the real-time tracker. This allows the system to adapt automatically to camera movements or environmental changes without human intervention. This optimized geometry feeds a multi-camera bird's-eye-view (BEV) tracker that fuses detections and unifies trajectories across zones. Crucially, by operating strictly on live anonymous metadata, our solution ensures a privacy-safe, zero-overhead, and resilient tracking pipeline that maintains global consistency in dynamic environments where static, recorded-video calibration is impossible.
David Voihanski, Mor Sinai, Ben Zion Bobrovsky
Juganu Or Yehuda, Israel · Tel Aviv University Israel
Reducing the number of cameras reduces the deployment cost but removes views that correct BEV responses stretched away from true pedestrian positions by projection and short score drops that can split tracks} in Bird's-Eye View (BEV) tracking. We introduce GRACE, a camera-efficient multi-view tracker with three components. Volumetric-Guided Fusion combines homography-based BEV features with features lifted through 3D space. Ray Conditioning exposes each camera's viewing direction to the fusion network. Its tracking component, BEV Track Recovery (BTR), uses low-confidence detections only to continue existing tracks. The same detections cannot start new tracks. With two WildTrack cameras, GRACE improves MOTA from 83.54 for TrackTacular, our baseline, to 91.07.