GRACE: Geometry- and Ray-Aware Camera-Efficient Multi-View Pedestrian Tracking

• Few-camera BEV tracking faces directional projection errors that displace responses and brief confidence drops that can split one track into two identities.Source: §1 · p. 1
• GRACE combines VGF, Ray Conditioning, and BTR: fusion balances homography and volumetric features, rays encode viewing direction, and BTR permits low-confidence detections to continue tracks but not start them.Source: §1 · p. 1
• TrackTacular is the primary baseline because it also performs detection and tracking in BEV.Source: §1 · p. 1
• 91.07 MOTA on two WildTrack cameras exceeds two-camera TrackTacular's 83.54 MOTA; GRACE also reaches 79.36 MOTA on MultiviewX.Source: §1 · p. 1
• The proposed geometry-aware fusion and recovery use only camera calibration and past observations.Source: §1 · p. 1

2.1. Problem Formulation

• With few cameras, fewer views can correct localization errors, while temporary confidence drops can split an existing track and assign a new identity.VGF and Ray Conditioning address projection errors; BTR addresses track splits.Source: §2.1 · p. 1

2. THE PROPOSED METHOD: GRACE

• 91.07 MOTA with two WildTrack cameras, versus 83.54 for two-camera TrackTacular; the seven-camera reference reaches 90.73 MOTA.The seven-camera reference uses VGF + Ray detections with the TrackTacular tracker.Source: Figure 1 · p. 2

2.2. Volumetric-Guided Fusion

• Volumetric-Guided Fusion mixes homography-projected features with features lifted through calibrated 3D space, using a learned gate and camera coverage.The coverage prior increases volumetric weight where few cameras can provide mutual correction; the learned term adapts the balance to observed features.Source: §2.2 · pp. 1–2

2.3. Ray Conditioning

• Ray Conditioning supplies each camera’s viewing direction at BEV locations, distinguishing locations with equal camera counts but different viewing directions.The two-channel direction map is computed from calibration and concatenated with each camera’s volumetric feature.Source: §2.3 · p. 2

2.4. BEV Track Recovery

• BEV Track Recovery matches low-confidence detections to existing tracks but bars them from initializing tracks, and predicts through missing detections for at most two frames under specified conditions.The low-confidence matching threshold is τ − δ with δ = 0.2; prediction is allowed when prior confidence is at least 0.4, and BTR remains online.Source: §2.4 · p. 2

3. EXPERIMENTAL SETUP

• All methods use WildTrack C1+C3 and MultiviewX C1+C6 on the datasets’ standard BEV evaluation regions.The same camera pair is used for every method on each dataset.Source: §3 · p. 2

3.1. Camera Configuration

• The two-camera setting is the minimum that enables multi-view fusion; the camera-count study also evaluates three to seven cameras.The study examines how the effects change as more views become available.Source: §3.1 · p. 2
• With two cameras, directional spreading is more visible in BEV responses, while additional views produce more concentrated responses.The WildTrack camera-count study reports BTR’s largest gain with two cameras.Source: Figure 3 · p. 3, Figure 6 · p. 4
• C1+C3 is the only WildTrack pair with complete BEV coverage; 44.98% of the grid is visible to both cameras.For MultiviewX C1+C6, at least one camera covers 95.62% of the evaluation area and both cameras see 44.21%.Source: §3.1 · p. 3, Figure 5 · p. 3
• WildTrack experiments fix C1+C3 across two-camera runs, follow TrackTacular preprocessing and ResNet-18 training, and share input processing, detection threshold, and evaluator across framework methods.The camera pair is selected using calibration before model evaluation.Source: §3.1 · p. 3

4.1. WildTrack Results

• With the same C1+C3 input, GRACE improves WildTrack results over TrackTacular from 82.53 to 86.38 MODA and from 83.54 to 91.07 MOTA.CaMuViD and RMCS are detection-only methods, so their tracking cells are omitted; the seven-camera block compares TrackTacular and GRACE.Source: §4.1 · p. 4

4.2. Camera-Count Study

• Across two to seven WildTrack cameras, VGF + Ray raises MODA from 86.38 to 92.58, versus 83.54 to 88.50 for TrackTacular.BTR adds 3.43 MOTA points with two cameras but at most 0.76 with three to seven; parameters remain fixed.Source: §4.2 · p. 4

4.3. MultiviewX Results

• With two cameras, GRACE reaches 79.36 MOTA on MultiviewX, 11.09 points above TrackTacular; MODA rises from 71.26 to 77.09 and IDF1 from 67.34 to 75.05.The two-camera comparison uses the same C1+C6 input for all evaluated methods.Source: §4.3 · p. 4, Table 2 · p. 4

4.4. Component Contributions

• On WildTrack C1+C3, VGF improves MODA by 2.52, IDF1 by 7.43, and MOTA by 3.05 over TrackTacular; Ray Conditioning then improves MODA by 1.33 and Recall by 1.54.The reported precision reduction is 0.19, described as small relative to the recall gain.Source: §4.4 · p. 4, Table 3 · p. 4
• BTR leaves detections unchanged but raises IDF1 by 2.83 and MOTA by 3.43 over VGF + Ray, separating temporal association gains from spatial-evidence changes.The ablation is on WildTrack C1+C3.Source: §4.4 · p. 4, Table 3 · p. 4
• At seven cameras, BTR changes MOTA from 90.73 to 91.04, compared with its 3.43-point gain at two cameras; overlapping views already reduce short detection gaps.Source: §4.4 · p. 4

4.5. Effect of Ray Conditioning

• With the same two WildTrack cameras, TrackTacular produces elongated BEV activations around pedestrians, while GRACE produces more compact activations.The comparison attributes reduced directional spreading in this setting to Ray Conditioning.Source: §4.5 · p. 4, Figure 7 · p. 4

• With two WildTrack cameras, GRACE reaches 91.07 MOTA, 1.36 points higher than TrackTacular with seven cameras.Source: §5 · pp. 4–5
• VGF and Ray Conditioning address projection errors, while BTR continues an existing track when detection confidence temporarily falls below the normal threshold.Source: §5 · p. 4
• On MultiviewX, GRACE improves MOTA by 11.09 points over TrackTacular using the same two cameras.Source: §5 · p. 5