Cross-View Direct Tracking Methods

• Cross-view direct tracking is a technique that directly integrates evidence from multiple cameras for real-time, robust object and pose tracking.
• It leverages geometric constructs like 3D/BEV representations and learned view-agnostic embeddings to unify detection and identity assignment.
• Methods utilize strategies such as Hungarian matching, transformer-based updates, and graph algorithms to mitigate occlusion and identity fragmentation.

Cross-view direct tracking denotes a family of multi-camera or multi-view tracking formulations in which association is performed directly across views and time, rather than by first solving independent single-view tracking and then applying a post-hoc fusion or re-identification stage. Across the literature, the term covers several related paradigms: direct 2D–3D association against maintained 3D track states for multi-human pose tracking (Chen et al., 2020), detection-free cross-view object tracking through shared 3D feature volumes and bird’s-eye-view supervision (Xu et al., 27 Feb 2025), direct Hungarian matching on learned view-agnostic embeddings without calibration (Xian et al., 10 May 2026), unified query-based multi-view multi-object tracking in arbitrary camera configurations (Li et al., 24 May 2025), and point-level or correspondence-level tracking across synchronized videos with explicit camera geometry and cross-view attention (Koo et al., 1 Dec 2025). A common premise is that cross-view evidence is used during the tracking step itself, so that occlusion, target loss, identity fragmentation, and drastic viewpoint change are addressed by multi-view coupling rather than by a downstream reconciliation stage.

1. Historical emergence and problem formulation

Early formulations of cross-view direct tracking were motivated by the observation that multi-view systems should exploit geometric and temporal consistency jointly. In multi-human 3D pose estimation, one line of work argued that previous methods “associate 2D poses among all pairs of views from scratch at every frame,” whereas a more efficient alternative is to “match the 2D inputs with 3D poses directly in 3-space” and “retain the 3D pose for each person and update them iteratively via the cross-view multi-human tracking” (Chen et al., 2020). In a closely related calibrated setting, “Part-Aware Measurement for Robust Multi-View Multi-Human 3D Pose Estimation and Tracking” describes a streaming framework in which “assigning a 2D pose to a 3D skeleton simultaneously enforces cross-view (all views assign to the same 3D ID) and temporal consistency” (Chu et al., 2021).

The object-tracking literature generalized this idea beyond human pose. “Generic Multiview Visual Tracking” defined multiview direct tracking with synchronized cameras as predicting all per-view boxes $B_t=\{b_t^c\}$ from multiview inputs $I_t=\{I_t^c\}$ while sharing an online model among cameras and recovering unreliable views through cross-camera trajectory prediction (Wu et al., 2019). More recent work on multi-view object tracking has reframed the problem around learned feature integration: MITracker “can track any object in video frames of arbitrary length from arbitrary viewpoints,” with the key step being direct prediction from refined cross-view tokens into a “lightweight CenterNet-style head” and explicitly “no external detector is called” (Xu et al., 27 Feb 2025).

In multi-object tracking, the same conceptual shift appears in methods that avoid single-view-first decomposition. MCBLT states that, unlike pipelines that “first track in each camera then ReID-link across views,” it “immediately lifts all views into BEV space” and “runs a single GNN on the aggregated detections and ReID features,” so “there is no separate ‘cross-view matching’ stage” (Wang et al., 2024). FusionTrack similarly presents “the first fully end-to-end Transformer-based multi-view multi-object-tracking framework that directly associates trajectories across an arbitrary number of (even non-overlapping) cameras without any external calibration or geometric priors” (Li et al., 24 May 2025).

This suggests a broad technical definition: cross-view direct tracking is less a single architecture than a design principle in which cross-view evidence is injected into state estimation, token refinement, graph association, or correspondence prediction before identities and trajectories are finalized.

2. Geometric representations and state spaces

A major axis of variation is the representation in which cross-view coupling is performed. In calibrated systems, several methods lift observations into a common 3D or BEV space. MITracker begins with per-view feature maps $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$ and projects candidate 3D points using camera intrinsics and extrinsics,

$\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$

then constructs a fused 3D feature volume

$$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$

which is compressed into a BEV representation $F'_{3D}\in\mathbb R^{C\times X\times Y}$ and supervised through a BEV focal loss $L_{bev}$ (Xu et al., 27 Feb 2025).

CC-3DT adopts a detection-level variant of the same principle. Each camera produces 3D detections

$$\mathcal{D}_t^{(c)}=\{d_i^{(c)}=(\mathbf b_i^{(c)},\mathbf f_i^{(c)},s_i^{(c)})\},$$

which are mapped into a common world frame through the calibrated transform

$$T_c^w= \begin{bmatrix} R_c^w & t_c^w\ \mathbf 0^\top & 1 \end{bmatrix}, \qquad X_i^w=R_c^w X_i^{(c)}+t_c^w,$$

and then fused by 3D NMS in BEV with $\tau_{\rm nms}=0.1$ before temporal association (Fischer et al., 2022). MCBLT similarly defines a world frame $I_t=\{I_t^c\}$0, projects reference points from BEV cells into each image, and uses spatial cross-attention to obtain a bird’s-eye feature map $I_t=\{I_t^c\}$1 in world coordinates (Wang et al., 2024).

Pose-tracking formulations maintain explicit 3D track states. In “Cross-View Tracking for Multi-Human 3D Pose Estimation at over 100 FPS,” a track is a 3D pose

$I_t=\{I_t^c\}$2

optionally with joint-wise velocities $I_t=\{I_t^c\}$3, and optimization balances reprojection cost with a motion prior,

$I_t=\{I_t^c\}$4

with $I_t=\{I_t^c\}$5–$I_t=\{I_t^c\}$6 (Chen et al., 2020). The part-aware method uses known $I_t=\{I_t^c\}$7 to project 3D joints by

$I_t=\{I_t^c\}$8

so that cross-view and temporal association are both mediated by maintained 3D skeletons (Chu et al., 2021).

Other formulations replace explicit geometry with learned latent spaces. CalibFree decomposes a shared ViT representation into view-agnostic features $I_t=\{I_t^c\}$9 and view-specific features $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$0, then average-pools $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$1 into a detection embedding $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$2 for cross-view matching (Xian et al., 10 May 2026). FusionTrack places all track queries $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$3 from all cameras into “the same $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$4-dimensional Transformer latent space,” adding globally consistent frame embeddings and camera-index embeddings so that cross-view alignment is learned rather than imposed geometrically (Li et al., 24 May 2025). VL-UniTrack extends the same idea to UAV-ground tracking by concatenating UAV and ground tokens inside a single shared encoder, producing a unified feature map $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$5 (Xu et al., 6 May 2026).

At the point and correspondence level, MV-TAP encodes camera rays through Plücker coordinates. For a point $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$6 it defines

$F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$7

and adds an MLP projection of $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$8 as a “camera embedding” to each token (Koo et al., 1 Dec 2025). This is distinct from explicit triangulation, but it still grounds cross-view tracking in calibrated multiview geometry.

3. Association mechanisms and direct update strategies

The defining operation in cross-view direct tracking is the association step. In geometric 2D–3D tracking, association is performed against persistent 3D states rather than among raw per-view detections. In the 2020 cross-view pose tracker, for each active track and new 2D pose detection, the affinity matrix combines a 2D term and a 3D term: $F_{2D}^k\in\mathbb R^{C\times H_s\times W_s}$9 The 2D affinity compares the new 2D joint with the last time the track was seen in that camera, scaled by plausible image-plane speed and an exponential decay; the 3D affinity predicts the joint location $\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$0, lifts the 2D detection to a 3D ray, and scores point-to-line distance. Matching is solved by the Hungarian algorithm (Chen et al., 2020).

The part-aware method uses a different 2D–3D affinity. For joint $\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$1,

$\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$2

and the full affinity averages positive per-joint affinities when the number of inlier joints exceeds $\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$3 (Chu et al., 2021). This formulation is explicitly “part-aware,” and before triangulation it applies a “joints filter” based on symmetric epipolar consistency and point-to-line distance from a predicted 3D joint to reject 2D outliers (Chu et al., 2021).

Cross-view direct tracking can also be realized without explicit state triangulation. MITracker compresses the fused BEV feature into a global multi-view token

$\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$4

concatenates this with each view’s unrefined token $\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$5, and refines it through transformer blocks using standard scaled-dot-product attention,

$\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$6

The refined tokens then “go directly into a lightweight CenterNet-style head to predict per-view heatmaps, box sizes, and offsets,” enabling “direct, detection-free, cross-view tracking” (Xu et al., 27 Feb 2025).

In calibration-free association, the central mechanism is direct matching in a learned invariant space. CalibFree computes a score between view-agnostic detection embeddings using either a Gaussian kernel on Euclidean distance,

$\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$7

or cosine similarity, forms a bipartite affinity matrix between active tracklets in different views, and applies the Hungarian algorithm. When a cross-view match conflicts with local IDs, “the global match wins,” and a global tracklet bank stores the most recent $\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$8 for each global identity (Xian et al., 10 May 2026).

Query-based multi-object trackers implement the same principle inside Transformer memory. FusionTrack stores all per-camera track queries in a Tracklet Memory Pool of length $\begin{pmatrix}u\v\1\end{pmatrix} =C_K\,[\,C_R\mid C_t\,]\, \begin{pmatrix}x\y\z\1\end{pmatrix},$9, applies a Multi-View ReID module over a temporal window $$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$0 to produce a unified ReID feature $$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$1, and then refines each query through an Object Update Module with both cross-frame and cross-view updates. Its cross-view attention takes the current query and attends to same-identity features in other cameras using a learnable view positional embedding (Li et al., 24 May 2025).

Graph-based approaches make association a global optimization problem over fused detections. MCBLT constructs a graph whose nodes are 3D detections with BEV position and aggregated appearance, passes messages by

$$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$2

predicts edge association scores $$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$3, and infers tracks through network flow or greedy linking (Wang et al., 2024). ModTrack implements a fully analytical alternative: after each sensor yields a calibrated position-covariance pair $$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$4, cross-view clustering is based on the $$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$5 distance

$$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$6

fusion is precision-weighted, and identity assignment is closed-loop through an identity-informed GM-PHD filter (Iyer et al., 16 Mar 2026).

4. Architectural families

Several recurring architectural families can be distinguished.

3D/BEV feature-integration trackers operate by converting multi-view image evidence into a shared spatial representation before tracking. MITracker is the clearest example: it transforms 2D features into a 3D feature volume, compresses that volume into BEV, and feeds the resulting global geometry back into each view through cross-view attention (Xu et al., 27 Feb 2025). CC-3DT and MCBLT use fused BEV detections rather than dense 3D feature volumes, but they share the same ordering: multi-camera fusion first, temporal association second (Fischer et al., 2022, Wang et al., 2024).

Direct 2D–3D matching trackers maintain a 3D latent state and match new 2D evidence to it. The 2020 cross-view pose tracker is iterative and one-view-at-a-time, updating tracks in chronological order and using weighted linear triangulation from current and recent observations (Chen et al., 2020). The part-aware method adds explicit outlier rejection through epipolar and predicted-ray filtering, then triangulates by DLT and fills missing joints by motion prediction or Gaussian temporal smoothing (Chu et al., 2021).

Calibration-free representation-learning trackers replace calibrated geometry with view-invariant descriptors. CalibFree learns view-agnostic embeddings via feature separation, single-view distillation, and cross-view reconstruction, then performs direct cross-view matching by Hungarian assignment (Xian et al., 10 May 2026). FusionTrack eliminates explicit calibration, homography, and depth maps entirely, relying on shared Transformer weights, frame/view encodings, cross-view attention, and view-guided hierarchical clustering plus neighbor filtering (Li et al., 24 May 2025).

Point and correspondence trackers treat the tracked entity as a point or a correspondence field rather than a bounding box or object state. MV-TAP predicts 2D trajectories $$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$7 and visibility flags $$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$8, using a three-axis Transformer that alternates attention along time, space, and view axes; view-attention is the key mechanism by which “an occluded or ambiguous point in one view can be ‘rescued’ by a clearer view” (Koo et al., 1 Dec 2025). MVTrack4Gen uses multi-view point tracking differently: it routes correspondence-specialized DiT attention features into an auxiliary tracking head and supervises both predicted tracks and attention weights to improve geometric consistency in novel-view video generation (Lee et al., 24 Jun 2026).

Cross-view correspondence and dual-view trackers focus on paired or sparse-view settings. VL-UniTrack is “a fully unified framework enhanced by visual-language prompts” for UAV-ground tracking; its shared encoder, visual-language geometric prompting, prompt-guided cross-view adapter, and confidence-modulated mutual distillation define a direct cross-view tracker that jointly localizes in both frames (Xu et al., 6 May 2026). V$$F_{3D}(c,x,y,z) =\frac{1}{|\mathcal V_{x,y,z}|} \sum_{k\in\mathcal V_{x,y,z}} F_{2D}^k\bigl(c,u_k(x,y,z),v_k(x,y,z)\bigr) \in\mathbb R^{C\times X\times Y\times Z},$$9-SAM extends SAM2 to cross-view correspondence via a geometry-aware anchor prompt generator, a visual prompt generator, three experts, and a post-hoc cyclic consistency selector (Pan et al., 25 Nov 2025).

Domain-specific and motion-compensated trackers show that cross-view direct tracking can also be defined by targeted geometric priors. The gymnast-tracking system uses a “cascaded DA paradigm” that triangulates when detections from multiple compatible views are available and falls back to ray–plane intersection when only a single detection or “exactly two opposite-view detections” are available (Yang et al., 20 Nov 2025). HomView-MOT, although described for moving UAV video rather than simultaneous multi-camera systems, realizes a related “cross-view direct tracking” principle through fast homography estimation, homographic box matching, and view-centric identity learning (Ji et al., 2024). GMT takes an even weaker-assumption route by requiring neither object-specific models nor camera calibration; it shares a collaborative correlation filter among cameras and invokes a cross-camera trajectory prediction network when a view becomes unreliable (Wu et al., 2019).

5. Benchmarks, datasets, and quantitative performance

The growth of cross-view direct tracking has been tied to new datasets and benchmarks. MITracker introduced the MVTrack dataset with “234K high-quality annotated frames featuring 27 distinct objects across various scenes,” and reported state-of-the-art results on both MVTrack and GMTD (Xu et al., 27 Feb 2025). DIVOTrack introduced “fifteen distinct scenarios and 953 cross-view tracks,” with 75 videos from “3 moving cameras,” and established standard benchmarks for cross-view multi-object tracking together with the CrossMOT baseline (Hao et al., 2023). The 2020 cross-view pose-tracking work also proposed “a new large-scale multi-human dataset with 12 to 28 camera views” (Chen et al., 2020). MDMOT, collected by “mobile drone swarms across diverse real-world scenarios,” serves as the benchmark for FusionTrack in arbitrary multi-view environments (Li et al., 24 May 2025).

A compact comparison of representative reported results is given below.

Method	Setting	Reported result
MITracker (Xu et al., 27 Feb 2025)	MVTrack	AUC = 71.13%, Normalized Precision = 91.87%, Precision = 83.95%
MITracker (Xu et al., 27 Feb 2025)	GMTD zero-shot	AUC = 65.96%, $F'_{3D}\in\mathbb R^{C\times X\times Y}$0 = 87.05%, $F'_{3D}\in\mathbb R^{C\times X\times Y}$1 = 82.07%
MITracker (Xu et al., 27 Feb 2025)	Recovery after full occlusion	79.2% vs. 56.7% for the next best
Cross-view pose tracking (Chen et al., 2020)	Store layout1, 12 cameras	154 FPS total
Cross-view pose tracking (Chen et al., 2020)	Store layout2, 28 cameras	34 FPS
Part-aware measurement (Chu et al., 2021)	Campus	96.79% average PCP
Part-aware measurement (Chu et al., 2021)	Shelf	97.39% average PCP
CC-3DT (Fischer et al., 2022)	nuScenes val	AMOTA = 35.9%, AMOTP = 1.361 m, IDS = 2152
CC-3DT (Fischer et al., 2022)	nuScenes test	AMOTA = 41.0%, AMOTP = 1.274 m, IDS = 3334
MCBLT (Wang et al., 2024)	AICity'24	81.22 HOTA
MCBLT (Wang et al., 2024)	WildTrack	95.6 IDF1
CalibFree (Xian et al., 10 May 2026)	MMP-MvMHAT overall	A = 58.4, F = 64.7 with YOLOX detections
FusionTrack (Li et al., 24 May 2025)	MDMOT single-view	MOTA = 88.13%, IDF1 = 92.04%, HOTA = 84.42%
FusionTrack (Li et al., 24 May 2025)	MDMOT multi-view	CVMA / CVIDF1 = 80.8% / 75.2%
MV-TAP (Koo et al., 1 Dec 2025)	DexYCB, 8 views	AJ 44.2% vs 41.5%, $F'_{3D}\in\mathbb R^{C\times X\times Y}$2 61.9% vs 59.6%, OA 78.3% vs 76.4%
ModTrack (Iyer et al., 16 Mar 2026)	WildTrack	IDF1 = 95.5, MOTA = 91.4, MOTP = 87.2, GOSPA = 1.20

Several results emphasize recovery and robustness rather than only steady-state accuracy. MITracker’s recovery experiment measures “re-acquiring the target within 10 frames after full occlusion” and reports 79.2% versus 56.7% for the next best method (Xu et al., 27 Feb 2025). On Shelf, replacing a single-view SORT baseline with cross-view tracking raised Camera 4 MOTA from 68.8% to 97.6%, IDF1 from 41.8% to 98.8%, and reduced ID switches from 3 to 0 (Chen et al., 2020). The gymnast-tracking system reported, under the hardest setting with only two opposite-view detections, “ID-switches reduced by 80–90 %,” “3D center AED reduced by ≈ 90 %,” and buffered 2D pose-box failure reduced from “~3 % to ~0.2 %” while running at “~425 FPS for tracking only” (Yang et al., 20 Nov 2025).

Dataset design also shapes the field’s assumptions. DIVOTrack explicitly addresses the lack of “real-world scenarios,” “diverse scenes,” “moving cameras,” and “standard benchmarks” in earlier cross-view tracking datasets (Hao et al., 2023). MVTrack is positioned as a response to the lack of comprehensive datasets for multi-view object tracking (Xu et al., 27 Feb 2025). This suggests that benchmark construction has been integral to defining what counts as successful cross-view direct tracking.

6. Technical challenges, misconceptions, and open directions

A recurring misconception is that cross-view direct tracking necessarily requires precise calibration and explicit 3D geometry. The literature does not support that claim. Calibrated methods such as MITracker, CC-3DT, MCBLT, MV-TAP, and the 2D–3D pose trackers rely centrally on known intrinsics, extrinsics, BEV projection, or ray geometry (Xu et al., 27 Feb 2025, Fischer et al., 2022, Wang et al., 2024, Koo et al., 1 Dec 2025, Chen et al., 2020). However, GMT requires “neither specific object model nor camera calibration,” instead learning an implicit dynamic mapping between cameras via its trajectory prediction network (Wu et al., 2019). CalibFree states that it “does not need any calibration or manual labeling” and performs cross-view association through learned view-agnostic embeddings (Xian et al., 10 May 2026). FusionTrack likewise uses “no external calibration or geometric priors” (Li et al., 24 May 2025). The field therefore contains both geometry-explicit and geometry-free variants.

Another misconception is that “direct” implies detector-free processing. MITracker indeed states that “no external detector is called” (Xu et al., 27 Feb 2025), but most multi-object and pose-tracking systems still begin from detections, 2D poses, or RoI crops. Cross-view pose tracking consumes detected 2D poses (Chen et al., 2020, Chu et al., 2021), CalibFree uses a person detector and RoI crops (Xian et al., 10 May 2026), ModTrack confines learning to “Detection and Feature Extraction” before performing analytical fusion and tracking (Iyer et al., 16 Mar 2026), and CrossMOT explicitly unifies detection, single-view association, and cross-view matching (Hao et al., 2023). In this literature, “direct” more commonly means that cross-view reasoning is integrated into the tracking update or identity assignment rather than delegated to a late post-processing stage.

Occlusion handling is the principal technical driver behind direct cross-view designs. MITracker argues that each view’s head can “re-localize the target even under occlusion” because it has access to global 3D context (Xu et al., 27 Feb 2025). The 2020 pose-tracking method notes that by matching in 3-space against the predicted $F'_{3D}\in\mathbb R^{C\times X\times Y}$3, “a track can skip frames in one view and re-appear later, robust to short occlusions,” and that incremental triangulation remains faithful when some views lag by up to “~300 ms” (Chen et al., 2020). ModTrack formalizes the same issue through uncertainty propagation and an identity-informed GM-PHD filter, emphasizing “traceable uncertainty” and transfer “unchanged” across WildTrack, MultiviewX, and RadarScenes except for the perception front-end (Iyer et al., 16 Mar 2026).

Limitations are correspondingly diverse. Geometry-free systems may degrade when “appearance cues alone may be ambiguous” under motion blur or illumination change, or when very large inter-view baselines provide little appearance overlap (Li et al., 24 May 2025). GMT notes failure when the source-view trajectory lacks variation or when cameras undergo very different rapid motions, and falls back to constant-velocity coasting when no view is reliable (Wu et al., 2019). VL-UniTrack identifies model size and inference cost as obstacles for on-board UAV deployment and notes that long-term failure remains difficult when both views degrade simultaneously (Xu et al., 6 May 2026). The gymnast tracker illustrates a different limitation: standard triangulation becomes ill-conditioned in sparse-view cases, motivating domain knowledge in the form of a predefined vertical plane (Yang et al., 20 Nov 2025).

A plausible implication is that future work will continue to split along three research directions already visible in the literature. One direction emphasizes richer geometric fusion, as in 3D feature lifting, BEV supervision, and uncertainty-aware analytical filters (Xu et al., 27 Feb 2025, Iyer et al., 16 Mar 2026). A second direction emphasizes calibration-free or free-viewpoint identity learning through unified latent spaces, prompting, and cross-view Transformers (Xian et al., 10 May 2026, Li et al., 24 May 2025, Xu et al., 6 May 2026). A third direction expands the unit of tracking itself—from objects and skeletons to arbitrary points, correspondences, and segmentation prompts—so that cross-view direct tracking becomes a generic correspondence engine useful not only for tracking but also for 4D video generation and cross-view object correspondence (Koo et al., 1 Dec 2025, Lee et al., 24 Jun 2026, Pan et al., 25 Nov 2025).