SMTrack: Temporal Tracking Methods

• SMTrack is a designation for various tracking frameworks that model temporal state evolution, differing in methodology from sliding-mode control to learned state-space prediction.
• It underpins applications including robust control with finite-time convergence, improved motion prediction via SSMs, and efficient RGB-video tracking employing spiking neural dynamics.
• The diverse implementations highlight trade-offs between computational efficiency and appearance reliance, offering practical insights for real-time tracking in uncertain environments.

Searching arXiv for "SMTrack" and closely related entries to ground the article in current papers. SMTrack is not a single canonical method but a recurring designation across several tracking literatures. In the sources considered here, it denotes a sliding-mode tracking and model-following controller for uncertain linear time-invariant systems (Ganguly et al., 2020), a state-space-model-based motion-prediction formulation instantiated by TrackSSM for multi-object tracking (Hu et al., 2024), an end-to-end spiking neural network for RGB-video multi-object tracking (Zhong et al., 20 Aug 2025), and a State-aware Mamba tracker for visual tracking (Ma et al., 2 Feb 2026). The shared name masks substantial differences in objective, temporal representation, and computational regime.

1. Terminological scope and problem settings

The designation SMTrack appears in at least four technically distinct contexts in the material considered here. The common denominator is the centrality of temporal state evolution, but the underlying mathematical object called “state” differs sharply across formulations: a control-theoretic sliding variable, an SSM hidden state for trajectory prediction, an SNN membrane-and-spike dynamics, or a Mamba-style latent state for visual tracking.

Usage of SMTrack	Problem setting	Defining components
Sliding-Mode Tracking	Robust tracking and model following for uncertain LTI systems	Super-twisting control, sliding surface, HOSMO/STO (Ganguly et al., 2020)
TrackSSM instantiation	Motion prediction in MOT	Flow-SSM, Mamba encoder, S$^2$L training (Hu et al., 2024)
Spiking SMTrack	RGB-video MOT	SpikeYOLOX, Asa-NWDLoss, TrackTrack association (Zhong et al., 20 Aug 2025)
State-aware Mamba SMTrack	Single-object visual tracking	SASM blocks, hidden-state propagation, Siamese pipeline (Ma et al., 2 Feb 2026)

A common misconception is that SMTrack refers to one benchmark lineage or one architecture family. The literature summarized here shows instead that it is an overloaded label spanning robust control, tracking-by-detection motion modeling, spiking detection-and-association pipelines, and state-space temporal modeling for visual tracking.

2. SMTrack in robust control: sliding-mode tracking and model following

In the control-theoretic usage, SMTrack refers to a robust tracking and model following design for uncertain LTI systems with matched uncertainty and disturbance. The plant is written as

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$

with lumped disturbance

$$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$

and the reference model is

$$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$

A matching condition is assumed through constant matrices $G$ and $H$, after which the tracking error is defined as

$e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$

with $z:=x-Gx_r$. The state is then decomposed via a coordinate transform into $(\eta,\xi)$, and a sliding variable

$$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$

is constructed with $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$0 chosen so that $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$1 is Hurwitz (Ganguly et al., 2020).

Control is split as $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$2, where $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$3 is a super-twisting term designed to drive $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$4 in finite time. In vector form,

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$5

and, componentwise,

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$6

The stated role of $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$7 is to exceed the bound on $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$8, thereby enforcing second-order sliding.

For the MagLev implementation, only the ball position $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$9 is measured, so velocity $$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$0 is estimated. A super-twisting observer is first attempted, but for continuous control the design adopts a third-order higher-order sliding-mode observer: $$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$1

$$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$2

$$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$3

The summary states that $$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$4 in finite time under a known bound on $$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$5, after which $$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$6.

The reported MagLev model is

$$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$7

The reference model is third order with poles at $$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$8. Reported results include tracking RMS error on the order of a few tens of microns, perfect disturbance rejection under $$w(t,x)=R(t,x)x+E(t,x)u+\Gamma(t,x)f(t,x), \qquad \|w(t,x)\|\le \theta_M,$$9, finite-time estimation of $$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$0 for both STO and HOSMO, and smoother control and sliding behavior for the HOSMO-based variant. In this usage, SMTrack is a control methodology centered on sliding-surface design, finite-time convergence, and observer-assisted output feedback rather than a learned tracker.

3. SMTrack as state-space motion prediction in TrackSSM

In the TrackSSM formulation, the provided description instantiates SMTrack as a motion predictor for tracking-by-detection. The central object is a discrete-time state-space model for a single track,

$$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$1

with a linear-Gaussian special case

$$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$2

TrackSSM’s Flow-SSM makes the effective time step $$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$3, as well as $$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$4 and $$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$5, depend on a learned flow feature $$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$6. Base parameters $$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$7 and $$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$8 are held fixed, while

$$\dot x_r = A_r x_r,\qquad y_r = C_r x_r,\qquad \|x_r(t)\|\le L.$$9

The discretization is performed by a zero-order-hold rule,

$G$0

followed by the update

$G$1

(Hu et al., 2024).

The flow feature $G$2 is obtained from a short trajectory history containing position and velocity terms,

$G$3

which is linearly embedded and processed by a stack of naive Mamba blocks. This produces a compact representation of recent trajectory dynamics. The decoder then receives the last-seen box $G$4, applies cosine positional encoding and linear splitting into $G$5 and $G$6, updates the Flow-SSM hidden state, gates the predicted embedding through an MLP-plus-sigmoid transform of $G$7, and maps the result back to a box prediction. Cascading $G$8 such layers yields stepwise refinement of $G$9.

The associated Step-by-Step Linear training strategy decomposes the one-step mapping $H$0 into $H$1 equal substeps. Pseudo-labels are generated by linear interpolation,

$H$2

so that each decoder layer is trained to predict an intermediate target. The total loss combines Smooth-$H$3 box regression and generalized IoU supervision.

The reported evaluation couples TrackSSM to YOLOX-x/l/m/s detectors and uses ByteTrack’s association. On MOT17, ByteTrack + KF achieves HOTA $H$4, while ByteTrack + TrackSSM achieves HOTA $H$5. On DanceTrack, the comparison is $H$6 versus $H$7 HOTA, and on SportsMOT $H$8 versus $H$9 HOTA. Additional ablations state that the best historical length is around $e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$0–$e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$1 frames, the decoder depth optimum is $e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$2, and S$e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$3L adds $e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$4 HOTA and $e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$5 IDF1 on DanceTrack. In this usage, SMTrack denotes a learned motion model with SSM-based temporal compression rather than a detector, controller, or association-only module.

4. SMTrack as an end-to-end spiking MOT framework

A later usage defines SMTrack as “the first directly trained deep SNN framework for end-to-end multi-object tracking on standard RGB videos” (Zhong et al., 20 Aug 2025). This framework is a two-stage pipeline composed of a SpikeYOLOX detector and a TrackTrack-based identity association module. The detector includes a spiking backbone, neck, and decoupled spiking head. The backbone/neck uses channel-mixing meta-blocks built from SepConv and ChConv operators, while the head follows a three-scale decoupled design inspired by YOLOX, with separate classification and regression branches.

The spiking dynamics are based on the Integer LIF neuron. In discrete time,

$e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$6

where $e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$7 is membrane potential, $e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$8 is the leak constant, $e(t)=y(t)-y_r(t)=C(x-Gx_r)=Cz,$9 is the firing threshold, and $z:=x-Gx_r$0 is the integer spike count in virtual time. Surrogate gradients replace the derivative of the Heaviside spike function, and the training summary specifies a fast sigmoid surrogate

$z:=x-Gx_r$1

Detection training is modified by the adaptive, scale-aware Normalized Wasserstein Distance loss. A bounding box $z:=x-Gx_r$2 is modeled as a Gaussian

$z:=x-Gx_r$3

with second-order Wasserstein distance

$z:=x-Gx_r$4

The normalized similarity is

$z:=x-Gx_r$5

where

$z:=x-Gx_r$6

Because $z:=x-Gx_r$7 is recomputed per batch from average object size, the method is described as increasing sensitivity to small objects under scale and density variation.

Association is handled by TrackTrack in three stages: motion prediction by an NSA Kalman filter, Track-Perspective Association using an affinity

$z:=x-Gx_r$8

with Hungarian matching over high-confidence, low-confidence, and NMS-suppressed detections in cascade, and Track-Aware Initialization for unmatched detections. The training setup initializes from COCO-pretrained YOLOX-s, trains on MOT17/MOT20 for $z:=x-Gx_r$9 epochs and DanceTrack/BEE24 for $(\eta,\xi)$0 epochs, uses SGD with momentum $(\eta,\xi)$1, weight decay $(\eta,\xi)$2, and sets virtual timesteps $(\eta,\xi)$3 with I-LIF integer range $(\eta,\xi)$4.

The reported results compare favorably with leading ANN baselines. On DanceTrack, SMTrack obtains HOTA $(\eta,\xi)$5, MOTA $(\eta,\xi)$6, and IDF1 $(\eta,\xi)$7; on BEE24, HOTA $(\eta,\xi)$8, MOTA $(\eta,\xi)$9, and IDF1 $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$0; on MOT17, HOTA $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$1, MOTA $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$2, and IDF1 $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$3; and on MOT20, HOTA $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$4, MOTA $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$5, and IDF1 $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$6. The paper summary adds that SMTrack uses only $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$7 or fewer timesteps while matching or exceeding ANN baselines on HOTA and IDF1. In this formulation, SMTrack is a learned detector-association system whose distinctive contribution lies in directly trained spiking computation for conventional RGB video.

5. SMTrack as State-aware Mamba for visual tracking

Another recent usage defines SMTrack as State-aware Mamba for efficient temporal modeling in visual tracking (Ma et al., 2 Feb 2026). This framework is a Siamese-style single-object tracker built from Selective State-Aware Space Model blocks. During training, $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$8 target templates and one search region are sampled from video, split into non-overlapping $$\sigma = [-K\;\; I_m]\begin{pmatrix}\eta\ \xi\end{pmatrix}$$9 patches, and embedded into tokens

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$00

with

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$01

Spatial positional embeddings, temporal embeddings, and target masks are added before concatenating template and search tokens and sending them through $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$02 blocks. A fully convolutional box head predicts a classification score map, an offset map, and a size map, from which the box $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$03 is recovered.

The key architectural departure from standard selective SSMs is a state-aware timescale. Starting from the continuous-time model

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$04

the discretized update becomes

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$05

Whereas Mamba makes $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$06, $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$07, and $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$08 content-dependent but shares timescale across state channels, SMTrack introduces

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$09

which are combined as

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$10

The hidden-state update is then

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$11

After scanning a frame’s patches, the last hidden state $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$12 is passed through a two-layer MLP and fed to the next block. This design is described as capturing more diverse temporal cues through state-wise parameters.

A major runtime feature is hidden-state propagation at inference. The initial template is scanned once, and the hidden-state outputs of each SASM block are stored. Each subsequent search region is processed by injecting these compressed hidden states into the blocks rather than reprocessing raw templates. Every $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$13 frames, a fresh template may be cropped, rescanned, and appended to memory. The stated training complexity of each block is

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$14

which is linear in the number of patches $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$15, in contrast to self-attention’s $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$16. At inference, after the initial template scan, each new search region incurs

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$17

operations and is independent of the number of stored templates.

The training data are COCO, LaSOT, GOT-10k, and TrackingNet, using four templates plus one search region, random translation, horizontal flip, and brightness jitter. The loss combines focal loss, an $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$18 box term, and GIoU,

$$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$19

with $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$20 and $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$21. Optimization uses AdamW, with learning rates $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$22 for SASM and $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$23 for the head, cosine decay, $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$24 epochs, and batch size $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$25.

Reported benchmark results include SMTrack-S256 at GOT-10k AO $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$26, LaSOT AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$27, TrackingNet AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$28, UAV AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$29, $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$30 GFlops, and $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$31 FPS; SMTrack-M256 at GOT-10k AO $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$32, LaSOT AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$33, TrackingNet AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$34, UAV AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$35, $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$36 GFlops, and $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$37 FPS; and SMTrack-M384 at GOT-10k AO $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$38, LaSOT AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$39, TrackingNet AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$40, UAV AUC $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$41, $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$42 GFlops, and $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$43 FPS. The ablation summary reports $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$44 AUC for state-wise $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$45 over state-shared, $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$46 AUC for the hidden-state interaction MLP, $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$47 AUC for multi-frame uniform sampling from memory, and a performance plateau at $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$48 hidden channels.

6. Comparative interpretation and relation to adjacent tracking research

Across these usages, SMTrack consistently denotes a method organized around explicit temporal state evolution, but the state variable itself is domain-specific. In sliding-mode control, the operative state is the sliding variable $$\dot x = A x + B\bigl(u + w(t,x)\bigr), \qquad y = Cx,$$49 and its observer-reconstructed plant state (Ganguly et al., 2020). In TrackSSM, it is a learned SSM hidden state driven by trajectory flow features (Hu et al., 2024). In the spiking MOT framework, temporal structure is represented through I-LIF membrane dynamics across a small number of virtual timesteps (Zhong et al., 20 Aug 2025). In State-aware Mamba tracking, temporal information is compressed into blockwise hidden states that are propagated across frames (Ma et al., 2 Feb 2026).

The computational and methodological implications are correspondingly different. The control-theoretic formulation seeks finite-time convergence and disturbance rejection. TrackSSM seeks motion prediction that improves association in tracking-by-detection, especially on nonlinear motion datasets. The spiking formulation seeks end-to-end RGB-video MOT with low-timestep spike-driven computation. The Mamba-based formulation seeks long-range temporal interaction with linear complexity in the number of patches and low per-frame inference cost through hidden-state reuse.

A second recurring theme is reduced reliance on appearance when temporal dynamics are informative. This is explicit in the TrackSSM motion module and in the TrackTrack association used by spiking SMTrack. A plausible implication is that SMTrack-labeled methods often emerge where conventional appearance-heavy pipelines are either too costly or insufficiently robust. That tendency is also visible in adjacent tracking work: YOLOv8-SMOT emphasizes appearance-independent association by integrating motion direction maintenance with an adaptive similarity metric inside OC-SORT for small multi-object tracking from UAV views (Yu et al., 16 Jul 2025).

Taken together, the literature does not support treating SMTrack as one architecture family. It is better understood as a recurrent naming pattern attached to temporally grounded tracking systems whose central innovation lies in how state is modeled, updated, and exploited: through higher-order sliding dynamics, data-dependent state-space motion prediction, spiking neural computation, or state-aware Mamba propagation.