Eva-Tracker: Robust Multi-Modal Tracking

• Eva-Tracker is a robust tracking framework that integrates aerial trajectory planning and event-based feature tracking to ensure continuous target visibility and collision avoidance.
• It employs an innovative FoV-ESDF representation to accelerate path optimization while managing occlusions effectively in dynamic environments.
• It also adapts to diverse sensor modalities with data-driven and cross-modal extensions, enhancing tracking performance under challenging conditions.

Eva-Tracker most directly denotes the ESDF-update-free, visibility-aware trajectory planning framework for robust aerial target tracking that was introduced to maintain continuous target visibility, avoid occlusions and collisions, and support target reacquisition without repeated Euclidean Signed Distance Field updates (Lin et al., 13 Feb 2026). The designation is also used in other cited works for distinct event-camera tracking systems: a data-driven feature tracker operating on grayscale templates and event patches (Messikommer et al., 2022), and, in a query-specific gloss, the cross-modal Gaussian-splatting event pose tracker GS-EVT (Liu et al., 2024). This suggests that the term is context-dependent across aerial planning, event-based feature tracking, and event-based camera tracking.

1. Name, scope, and referents

The cited literature uses “Eva-Tracker” in more than one technical sense. The formally titled use is the aerial tracking planner in “Eva-Tracker: ESDF-update-free, Visibility-aware Planning with Target Reacquisition for Robust Aerial Tracking” (Lin et al., 13 Feb 2026). A separate summary applies the same label to the event-camera feature tracker introduced in “Data-driven Feature Tracking for Event Cameras” (Messikommer et al., 2022). Another summary states that GS-EVT is “also referred to as Eva-Tracker in the query,” even though the paper title is “GS-EVT: Cross-Modal Event Camera Tracking based on Gaussian Splatting” (Liu et al., 2024).

Usage	Core task	arXiv id
Eva-Tracker	Visibility-aware aerial tracking planner	(Lin et al., 13 Feb 2026)
Eva-Tracker	Data-driven feature tracking for event cameras	(Messikommer et al., 2022)
GS-EVT referred to as Eva-Tracker	Cross-modal 6-DoF event-camera tracking	(Liu et al., 2024)

In the narrowest and most canonical sense, Eva-Tracker is the aerial planning framework of (Lin et al., 13 Feb 2026). In a broader bibliographic sense, the name has been attached to multiple systems that share a tracking objective but differ in sensing modality, state representation, and optimization structure.

2. Aerial tracking formulation and system structure

In its formal 2026 usage, Eva-Tracker addresses aerial target tracking with a drone that must continuously observe a moving target while avoiding collisions and recovering quickly when the target is temporarily lost (Lin et al., 13 Feb 2026). The paper identifies a bottleneck in conventional visibility-aware tracking: repeated ESDF updates introduce considerable computational overhead, even though the planner mainly needs to know whether obstacles enter the camera’s field of view and whether the target remains visible at an appropriate observation distance.

The framework replaces repeated visibility-field updates with a precomputed Field of View ESDF, or FoV-ESDF. Rather than storing obstacle distance over the whole environment, FoV-ESDF is tied to the camera geometry and is used for rapid visibility evaluation without requiring updates. The pipeline has three tightly coupled stages: target trajectory prediction, visibility-aware initial path generation, and FoV-ESDF-guided trajectory optimization (Lin et al., 13 Feb 2026).

The target trajectory predictor fits a smooth Bézier trajectory $\boldsymbol{\rho}(t)$ over a time window $[-T,T]$ from a history of observed target positions $\mathbf{h}_0,\dots,\mathbf{h}_m$ by minimizing the integral of squared second derivative subject to interpolation constraints:

$$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$

The paper states that this is formulated as an equality-constrained quadratic program with a closed-form solution (Lin et al., 13 Feb 2026).

3. Visibility-aware path generation and target reacquisition

Eva-Tracker’s initial path generation is explicitly recovery-capable. For each planning step, the method selects a candidate observation point along the ray from the predicted target position toward the tracker at a preferred distance $d$:

$$\mathbf{c}_k=\boldsymbol{\rho}(k\tau)-d\frac{\mathbf{\Pi}_2(\boldsymbol{\rho}(k\tau)-\mathbf{p}_{k-1})}{\|\mathbf{\Pi}_2(\boldsymbol{\rho}(k\tau)-\mathbf{p}_{k-1})\|},$$

where $$\mathbf{\Pi}_2=(\mathbf{e}_1,\mathbf{e}_2,\mathbf{0})$$ extracts the horizontal component (Lin et al., 13 Feb 2026).

If the segment from $\mathbf{c}_k$ to the predicted target is not occluded, $\mathbf{c}_k$ becomes the next waypoint. If it is occluded, the planner searches along a circular arc of radius $d$ around the target until it finds the nearest occlusion-free point, which becomes the waypoint $[-T,T]$0. The paper emphasizes that this yields a feasible, visibility-consistent initial path quickly, without global search.

The reacquisition mechanism is integrated into this same path generator rather than implemented as a separate recovery mode. Even when the target disappears, the planner continues generating waypoints around the predicted target position at the desired observation distance, so that once the target re-enters the field of view, tracking resumes smoothly (Lin et al., 13 Feb 2026). A plausible implication is that the system treats temporary loss primarily as a prediction-and-visibility problem rather than as a discrete mode switch.

4. FoV-ESDF representation and differentiable optimization

The central representation is the FoV-ESDF, defined over the camera’s quadrilateral-pyramid field of view:

$[-T,T]$1

where $[-T,T]$2 and $[-T,T]$3 are the horizontal and vertical FoV angles and $[-T,T]$4 is chosen as the optimal observation distance $[-T,T]$5 (Lin et al., 13 Feb 2026). Unlike a standard ESDF, FoV-ESDF stores the distance from points inside the camera FoV to the nearest FoV boundary and sets the value to zero outside the FoV. Dynamic obstacles are still handled by transforming obstacle points into the tracker frame at each step, but the field itself is not recomputed.

The planner parameterizes waypoints and yaw angles as a MINCO trajectory and optimizes them with L-BFGS under a composite objective:

$[-T,T]$6

Here, $[-T,T]$7 is the visibility term, $[-T,T]$8 is the energy or smoothness term, and $[-T,T]$9 enforces collision and kinematic constraints (Lin et al., 13 Feb 2026).

The visibility cost is decomposed as

$\mathbf{h}_0,\dots,\mathbf{h}_m$0

with differentiable components for occlusion, observation distance, and yaw alignment. The observation-distance term drives the target toward the maximum FoV-ESDF value at $\mathbf{h}_0,\dots,\mathbf{h}_m$1 in the tracker frame, while the yaw term encourages the drone to face the target directly (Lin et al., 13 Feb 2026). The paper also states that RC-ESDF is used only inside the safety penalty, so the method is “ESDF-update-free” for visibility planning, not necessarily for all collision logic.

5. Reported performance, computation, deployment, and limitations

The simulation benchmark in (Lin et al., 13 Feb 2026) uses the same FoV $\mathbf{h}_0,\dots,\mathbf{h}_m$2, preferred distance $\mathbf{h}_0,\dots,\mathbf{h}_m$3 m, and limits of 2 m/s linear velocity and 1.5 rad/s yaw rate for all compared methods. The environment contains 100 random obstacles and a 200-second target trajectory with sharp turns and cluttered segments. Metrics are tracking distance (TD), yaw angle error (AE), occlusion rate (OR), and failure rate (FR), with failure defined as the target leaving the FoV or becoming occluded.

Eva-Tracker is reported to achieve the best average tracking distance, $\mathbf{h}_0,\dots,\mathbf{h}_m$4 m, a yaw angle error of $\mathbf{h}_0,\dots,\mathbf{h}_m$5 rad, the lowest occlusion rate at 4.36%, and the lowest failure rate at 6.75% (Lin et al., 13 Feb 2026). The comparison values reported in the paper are SF-Tracker with TD $\mathbf{h}_0,\dots,\mathbf{h}_m$6, AE $\mathbf{h}_0,\dots,\mathbf{h}_m$7, OR 27.50%, FR 40.58%; Vis-Planner with TD $\mathbf{h}_0,\dots,\mathbf{h}_m$8, AE $\mathbf{h}_0,\dots,\mathbf{h}_m$9, OR 17.32%, FR 28.50%; and Elastic Tracker with TD $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$0, AE $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$1, OR 8.71%, FR 12.27%.

Computation time is another central result. The initial path generator takes only $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$2 ms on average, the optimization stage takes $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$3 ms, and total planning time is $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$4 ms (Lin et al., 13 Feb 2026). The paper contrasts this with 36.69 ms for Vis-Planner and 37.52 ms for SF-Tracker, while noting that Elastic Tracker reaches 7.12 ms total but does not provide full 3D occlusion reasoning.

The real-world system is deployed on a quadrotor with a Livox Mid-360 LiDAR and an onboard NVIDIA Jetson Orin NX, using Faster-LIO for localization, YOLOv11 for person detection, and MonoLoco for monocular 3D pose estimation (Lin et al., 13 Feb 2026). In outdoor tests with trees, the drone navigates around trunks while keeping the target visible and at a stable distance. In indoor tests, the target passes through a doorway that closes and blocks the view; the drone uses historical observations to predict where the target will go, generates waypoints on the far side of the occlusion, and reacquires the target once it reappears.

The limitations are explicit. The reacquisition mechanism works best for short-term target loss; if occlusion lasts long enough and the target’s true motion diverges from the prediction, recovery becomes difficult. FoV-ESDF also assumes a fixed camera field of view, so it does not directly handle zoom cameras or gimbal systems with changing FoV (Lin et al., 13 Feb 2026).

6. Other systems called “Eva-Tracker” in event-camera literature

A separate usage applies the name to the event-camera feature tracker described in “Data-driven Feature Tracking for Event Cameras” (Messikommer et al., 2022). There, the tracker follows a hybrid frame-event pipeline: a feature is first detected in a grayscale reference frame as a template patch $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$5, then event patches $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$6 are extracted around the previous feature location from the asynchronous event stream, and the network predicts the relative feature displacement $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$7. The architecture combines separate FPN-based patch encoders, a correlation map, a ConvLSTM block, and a frame attention module that performs self-attention across tracks in the same image. The method is trained on Multiflow, transfers zero-shot from synthetic to real data, and is reported to improve relative feature age by up to 120%, with the performance gap increased to 130% after pose-based adaptation (Messikommer et al., 2022).

A third usage appears in the description of GS-EVT, which is stated to be “also referred to as Eva-Tracker in the query” (Liu et al., 2024). GS-EVT is a cross-modal 6-DoF event-camera tracking method that aligns integrated event images with differential renderings from a Gaussian Splatting map built from RGB frames. Its core modeling choice is a reference pose plus first-order dynamics, with rendered intensity-change images

$$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$8

compared to accumulated event images in a staggered coarse-to-fine optimization over pose and velocity. On VECtor, it is reported to be 2–10× more accurate than EVPT, and on self-collected sequences it outperforms both EVPT and EVT in difficult scenarios such as severe occlusion, complex texture, and highly reflective surfaces (Liu et al., 2024).

Taken together, these uses of the same name span visibility-aware aerial planning, event-driven feature propagation, and cross-modal event-camera pose estimation. The common thread is robust tracking under difficult sensing conditions, but the underlying state variables differ fundamentally: trajectory and visibility constraints in (Lin et al., 13 Feb 2026), feature displacement in (Messikommer et al., 2022), and camera pose in $$\begin{aligned} \underset{\boldsymbol{\varrho}}{\min}\ & \int_{-T}^{T} \|\boldsymbol{\rho}^{(2)}(t)\|^2 \, dt, \ \text{s.t.}\ & \boldsymbol{\rho}(i\tau - T)=\mathbf{h}_i,\quad i\in\{0,\dots,m\}. \end{aligned}$$9 in (Liu et al., 2024).