AMOT: Disambiguation and Tracking Applications

• AMOT is an overloaded term in computer vision that denotes both active multi-object tracking via multi-camera systems and appearance-guided UAV tracking methods.
• In active multi-object tracking, coordinated cameras use methods like YOLO detection and Double Q-learning to optimize target coverage, achieving around 71.9% performance compared to fixed baselines.
• The UAV-based AMOT approach integrates appearance-motion consistency with motion-aware track continuation, yielding state-of-the-art IDF1 and MOTA scores on benchmarks such as VisDrone2019 and UAVDT.

Searching arXiv for papers and acronym disambiguation around “AMOT”. AMOT is an overloaded acronym rather than a single stable term in arXiv usage. In computer vision, it is used explicitly for Active Multi-Object Tracking, a multi-camera control problem in which cameras move, rotate, and zoom to maximize target coverage (Fang et al., 2022), and for Appearance-guided Motion modeling for robust Multi-Object Tracking, a UAV-oriented tracker that couples appearance and motion through bi-directional spatial consistency (Ma et al., 3 Aug 2025). Closely related literature also connects AMOT queries to AMOTA, the recall-averaged tracking metric used in nuScenes-style 3D multi-object tracking benchmarks (Wang et al., 2022), and to broader motion-association formulations in online MOT (Yin et al., 2020). By contrast, the atomic-physics term most likely to be confused with AMOT in one relevant paper is AF-MOT, not AMOT, and denotes an alternating-frequency magneto-optical trap (Wiegand et al., 2019).

1. Terminological scope and disambiguation

The literature represented here uses “AMOT” in several technically distinct ways. The most explicit uses are task-level and method-level definitions in multi-object tracking, while several near matches arise from benchmark nomenclature or orthographically similar acronyms. This suggests that AMOT should be read contextually rather than assumed to denote one canonical framework.

Usage	Domain	Source
Active Multi-Object Tracking	Multi-camera active vision and MARL	(Fang et al., 2022)
Appearance-guided Motion modeling for robust Multi-Object Tracking	UAV MOT	(Ma et al., 3 Aug 2025)
AMOTA / AMOTP often retrieved in “AMOT” searches	3D MOT evaluation	(Wang et al., 2022)
AF-MOT, not AMOT	Atomic physics, magneto-optical trapping	(Wiegand et al., 2019)

A further source of ambiguity is that several papers relevant to AMOT-like searches do not introduce a method named AMOT. “CAMO-MOT” is a camera–LiDAR fusion 3D MOT framework that explicitly defines AMOTA and sAMOTA and reports a nuScenes leaderboard score of 75.3% AMOTA (Wang et al., 2022). “IMM-MOT” improves nuScenes AMOTA through interacting multiple-model filtering and reports 73.8% on nuScenes validation (Liu et al., 13 Feb 2025). “AutoMoT” is a distinct vision-language-action model for end-to-end autonomous driving, not a multi-object tracker (Huang et al., 16 Mar 2026).

2. AMOT as Active Multi-Object Tracking

In its clearest explicit sense, AMOT denotes Active Multi-Object Tracking, defined as a setting in which multiple controllable cameras are coordinated by a centralized system to maximize target coverage in a shared environment (Fang et al., 2022). The task differs from passive MOT because camera pose is part of the control loop, and it differs from ordinary camera control because the objective is collaborative, multi-target, and partially observable. The formalization is a multi-agent POMDP

$$\left\langle S, A, T, R, \Omega, O, \gamma, I \right\rangle,$$

with joint observation

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$

and joint action

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$

The optimization target is

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$

The central technical device is coordinate-aligned multi-camera collaboration. Each camera detects targets with YOLOv4-Tiny, then inverse-projects detections from image coordinates into a common world-coordinate system under a ground-plane assumption $z=0$. With intrinsic matrix

$$K= \begin{bmatrix} f_x & 0 & u_0\ 0 & f_y & v_0\ 0 & 0 & 1 \end{bmatrix},$$

the paper gives the inverse mapping

$$\begin{bmatrix} x\ y\ 1 \end{bmatrix} = Z_c \begin{bmatrix} r_1 & r_2 & T_{3\times1} \end{bmatrix}^{-1} K^{-1} \begin{bmatrix} u\ v\ 1 \end{bmatrix}.$$

The image point $(u,v)$ is taken as the midpoint of the bottom edge of the bounding box, so that the recovered $(x,y)$ approximates the target’s ground contact location.

The policy is learned with a centralized Double Q-learning architecture. Each camera is an agent with a discrete 27-way action space,

$a = [a_m, a_r, a_z],$

corresponding to translation, rotation, and zoom, each with three choices. The network encodes per-agent observations, concatenates them with a one-hot agent identity code and previous action, and outputs branched Q-values for the three action dimensions. The update is

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$0

with

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$1

Reward shaping is essential. The per-agent reward is

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$2

where the team reward is the coverage rate

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$3

and the individual reward decomposes into bounding-box, visibility, direction, and position terms. The visibility term

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$4

penalizes redundant overlap, while the direction term

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$5

encourages stable centering of covered targets.

Evaluation uses a custom 3D Unreal Engine environment, “Soccer Court,” with 22 human-like players and 6 mobile cameras. The main reported result is 71.88% coverage, outperforming the fixed-camera baseline by 8.9%; the table reports 71.9 ± 5.8\% for the full method, 63.0 ± 4.5\% for the baseline, 72.1 ± 5.0\% for an oracle-box variant, and 66.9 ± 5.8\% without inverse projection (Fang et al., 2022). Ablations show that removing the direction reward or bounding-box reward degrades coverage strongly, and that team-only reward underperforms the mixed team-plus-individual design. In this sense, AMOT is not merely tracking but active sensing under joint geometric state alignment and MARL-based control.

3. AMOT as Appearance-guided Motion modeling for robust Multi-Object Tracking

A second explicit use of the acronym is the UAV tracker AMOT, whose name expands to Appearance-guided Motion modeling for robust Multi-Object Tracking (Ma et al., 3 Aug 2025). This AMOT is not a control problem; it is a plug-and-play, training-free enhancement for JDE-style UAV MOT. The method is built on a FairMOT-style joint detection and embedding framework and a ByteTrack-like two-stage association strategy, and addresses the instability of motion-only and appearance-only affinity in UAV videos.

The main module is the Appearance-Motion Consistency (AMC) matrix. For each previous track embedding $$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$6, the tracker computes a dense response map over the current-frame ReID feature map,

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$7

and selects the max-response location

$$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$8

Symmetrically, each current detection embedding $$o_t = \langle o_{1,t}, o_{2,t}, \dots, o_{n,t}\rangle$$9 is projected backward onto the previous-frame feature map,

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$0

yielding

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$1

These appearance-guided locations are converted into forward and backward spatial distances,

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$2

and then into the AMC cost

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$3

Stage-1 association fuses AMC with appearance and IoU costs through

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$4

followed by Hungarian matching.

The second module is Motion-aware Track Continuation (MTC), which attempts to preserve tracks through missed detections. For unmatched tracks retained in a buffer

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$5

the tracker compares a Kalman-predicted center $$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$6 with an appearance-guided center $$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$7 derived from the dense response map

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$8

Reactivation occurs if

$$a_t = \langle a_{1,t}, a_{2,t}, \dots, a_{n,t}\rangle.$$9

and there is no significant overlap with current detections.

Empirically, this AMOT reports state-of-the-art results on three UAV benchmarks. On VisDrone2019, it achieves IDF1 = 61.4, MOTA = 46.0, and IDs = 1063; on UAVDT, IDF1 = 74.7, MOTA = 55.1, and IDs = 272; on VT-MOT-UAV, IDF1 = 52.7 and MOTA = 31.8 (Ma et al., 3 Aug 2025). The ablations attribute most of the ID-switch reduction to AMC, while MTC primarily improves continuity under missed detections. Runtime remains close to baseline, with 36.4 FPS versus 37.1 FPS for the underlying tracker. In this use, AMOT names a specific UAV tracking method centered on appearance-guided motion regularization.

4. AMOT, AMOTA, and 3D MOT benchmark usage

A substantial fraction of “AMOT” retrievals in computer vision correspond not to a method named AMOT but to the metric family AMOTA/AMOTP. “CAMO-MOT” provides the most explicit metric definitions in the data. It writes

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$0

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$1

and

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$2

with the important note that the official nuScenes “AMOTA” is in fact the scaled version $$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$3 (Wang et al., 2022). CAMO-MOT is a camera–LiDAR fusion framework with three modules—O2S-OAM, CS-MM, and M2-FAM—and reports 75.3% AMOTA on the nuScenes test set, as well as 23 IDS for Car and 137 IDS for Pedestrian on KITTI test.

“IMM-MOT” is a LiDAR-only tracking-by-detection framework that targets exactly the error sources that strongly affect AMOTA: inaccurate motion prediction, premature trajectory death, and poor score filtering (Liu et al., 13 Feb 2025). Its core contribution is an interacting multiple-model tracker over the model set

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$4

with fused prediction

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$5

and model-probability update

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$6

On nuScenes validation it achieves AMOTA 73.8%, AMOTP 51.6, IDS 326, FP 13433, and FN 15658. The ablation table shows that IMM contributes the largest isolated AMOTA gain, from 73.08 to 73.59.

Other autonomous-driving trackers in the same retrieval neighborhood reinforce the same point. A two-stage 3D data-association method reports 0.587 AMOTA on nuScenes validation by partitioning tracklets into high- and low-confidence sets before local and global assignment (Dao et al., 2021). MotionTrack, an end-to-end transformer-based LiDAR-camera fusion baseline, reports AMOTA 0.55 on nuScenes test for its best VoxelNet fusion variant (Zhang et al., 2023). In practical retrieval terms, AMOT is therefore frequently entangled with AMOTA-centric autonomous-driving tracking literature, even when the acronym does not denote the method name itself.

5. Joint motion–association modeling and identity-aware MOT

Beyond explicit AMOT usage, several papers define a broader methodological region in which motion prediction and association are no longer treated as isolated modules. A representative example is UMA, which unifies object motion and affinity learning in a single online MOT network by combining SiamFC-style single-object tracking with metric learning inside a shared triplet architecture (Yin et al., 2020). The response map is formed as

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$7

while identity-aware embeddings are trained with an N-pair loss

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$8

On MOT17 test, UMA reports MOTA 53.1, IDF1 54.4, and IDS 2,251, with a model size of about 30M and runtime around 5.0 fps.

Identity-centric association is pushed further by SAMOT, a switcher-aware online tracker that explicitly models confusing nearby rivals, or “switchers,” as the proximate cause of identity errors (Feng et al., 2020). Its Spatial Conflict Graph selects a subgraph

$$\mathbb{E}\left[\sum_{i=1}^n \sum_{t=1}^{T} R_{i,t}\right].$$9

with

$z=0$0

while Switcher-Aware Association conditions each track–detection decision on a candidate temporal switcher mined from a heat map. On MOT17 test it reports MOTA 52.4%, IDF1 56.3%, and IDS 2166. The same paper also proposes SAIDF, an identity-focused measure designed to be more sensitive to identity fragmentation than MOTA and less segment-selective than IDF1.

A related, though not association-centric, contribution is EMAP, an ego-motion-aware target prediction module for moving-camera MOT (Mahdian et al., 2024). It modifies the Kalman prediction step by adding explicit camera-motion control terms,

$z=0$1

thereby separating object motion from ego-induced image motion. On KITTI, EMAP reduces the identity switches of OC-SORT from 53.24 to 14.10 and raises HOTA from 67.74 to 70.21. Taken together, these papers support a broader interpretation in which AMOT-like queries often target motion–association coupling, identity preservation, and autonomous-scene robustness, rather than one single acronymic entity.

6. Non-tracking lookalikes and cross-domain ambiguity

Not every near match to AMOT belongs to multi-object tracking. In atomic physics, the paper “A single-laser alternating-frequency magneto-optical trap” explicitly states that it does not use the acronym AMOT; the correct term is AF-MOT, for alternating-frequency magneto-optical trap (Wiegand et al., 2019). The system uses a single agile laser that sequentially switches between cooling and repumping transitions by tuning the injection current of a micro-integrated extended-cavity diode laser. For $z=0$2, it demonstrates optical frequency jumps of up to 6.6 GHz with tuning time around 15 $z=0$3s, repetition rate up to 7.6 kHz, maximum atom number about $z=0$4, and cloud temperature of about 500 $z=0$5K. The paper is relevant to AMOT searches only because AF-MOT may be informally conflated with AMOT; terminologically, however, it is a different acronym.

A separate ambiguity arises from AutoMoT, a vision-language-action architecture for end-to-end autonomous driving (Huang et al., 16 Mar 2026). AutoMoT is not an MOT system in the standard sense. It couples a frozen Qwen3-VL-4B understanding expert to a roughly 1.6B-parameter action expert through layer-wise joint attention sharing,

$z=0$6

$z=0$7

and uses asynchronous fast–slow inference. It reports Driving Score 87.34 and Success Rate 70.00% on Bench2Drive, and 0.32 average L2 with 0.07 average collision rate on nuScenes open-loop planning. Its inclusion in AMOT-like searches reflects acronymic proximity, not conceptual identity.

Across these usages, the most stable conclusion is that AMOT is a context-dependent label. In the narrowest explicit sense it refers either to Active Multi-Object Tracking (Fang et al., 2022) or to the UAV tracker Appearance-guided Motion modeling for robust Multi-Object Tracking (Ma et al., 3 Aug 2025). In broader retrieval practice it often points toward AMOTA-centric 3D tracking, motion–association unification, or unrelated lookalike acronyms such as AF-MOT and AutoMoT.