VFEO: Enhanced Obstacle Estimation

• VFEO is a framework that forms robust obstacle-centric variables from complementary sensing channels and optimizes sensor geometry to reduce estimation errors.
• It integrates passive vision methods (Mask R-CNN, stereo disparity, optical flow) with cooperative UAV sensing (DM-RS based range and Doppler) for enhanced obstacle tracking.
• Empirical evaluations demonstrate significant CRLB reductions and improved position accuracy in both road-scene perception and coordinated multi-UAV configurations.

Searching arXiv for the specified papers and closely related context. arxiv_search(query="(Rateke et al., 2020)", max_results=5, sort_by="relevance") arxiv_search(query="(Wang et al., 29 Aug 2025)", max_results=5, sort_by="relevance") arxiv_search(query="\"Variable Formation Enhanced Obstacle Position Estimation\" OR VFEO", max_results=10, sort_by="relevance") Variable Formation Enhanced Obstacle Position Estimation (VFEO) denotes an obstacle-state estimation paradigm in which variables derived from complementary sensing channels are explicitly formed, fused, and, when necessary, supported by geometry adaptation to improve position inference. In the arXiv literature represented here, the term appears explicitly in a 2025 cooperative sensing framework for multi-UAV path-following and obstacle avoidance, where UAV formation geometry is reconfigured to reduce the position Cramér–Rao lower bound (CRLB) (Wang et al., 29 Aug 2025). A closely related usage is applied as a design and implementation blueprint for a 2020 road-obstacle perception stack that combines Mask R-CNN instance segmentation, stereo disparity maps, and Gunnar Farnebäck optical flow to extract class, position, depth, and motion from passive vision alone (Rateke et al., 2020). This suggests that VFEO is best understood as a family of estimation strategies centered on structured variable formation, object-centric fusion, and uncertainty-aware decision support.

1. Nomenclature and conceptual scope

In the most literal sense, VFEO is the cooperative sensing algorithm proposed for a multi-UAV system using integrated sensing and communication (ISAC) signals, information-level fusion, and variable formation through reconfiguration of UAV positions on a formation circle (Wang et al., 29 Aug 2025). In a broader but explicitly motivated sense, the 2020 road-obstacle work underpins a VFEO-style methodology because it forms per-obstacle variables from instance masks, disparity statistics, and optical-flow statistics, then uses those variables for enhanced position estimation and threat-oriented interpretation (Rateke et al., 2020).

Formulation	Sensing basis	Core outputs
Multi-UAV VFEO	5G OFDM DM-RS ISAC, range and radial velocity fusion	obstacle position, obstacle velocity, CRLB-driven formation reconfiguration
Passive-vision VFEO blueprint	Mask R-CNN, stereo disparity, dense optical flow	class, 2D and 3D position, depth, motion, TTC, threat score

A common conceptual thread is the replacement of raw pixel-level or sensor-level observations with obstacle-centric variables that are more suitable for fusion. In the road context, these variables are formed inside each instance mask. In the UAV context, they are formed from per-UAV range and Doppler estimates and then reconditioned through formation changes. A plausible implication is that the phrase “variable formation” names two distinct but compatible operations: the construction of robust estimation variables from measurements, and the reconfiguration of sensor geometry so that those variables become more informative.

2. Passive-vision object-centric estimation

The passive-vision formulation combines three components: object instance segmentation, stereo disparity, and dense optical flow (Rateke et al., 2020). The detector is Mask R-CNN with an Inception backbone, pre-trained on MS COCO and applied to the left camera image. For each detection $i$, the system outputs a class label $c_i$, detection confidence $s_i$, a 2D bounding box $B_i$, and a pixel-wise mask $M_i$. Stereo disparity is computed from rectified left/right image pairs with disparity defined as

$D(u,v)=x_l-x_r,$

yielding a disparity map with per-pixel disparity $d_t(u,v)$. Dense motion is estimated between consecutive left images using Gunnar Farnebäck optical flow, with per-pixel motion $(u_t(u,v),v_t(u,v))$.

The geometric reconstruction model is the stereo pinhole model. For a pixel with disparity $d>0$,

$Z=\frac{fB}{d},$

where $c_i$0 is the focal length in pixels and $c_i$1 is the baseline in meters. The corresponding 3D point in the left-camera frame is

$c_i$2

with

$c_i$3

If vehicle-frame coordinates are required, the transform is

$c_i$4

Motion is derived from optical flow and disparity dynamics. The optical-flow constraint is

$c_i$5

Within each obstacle mask, the method computes robust median flow $c_i$6 and uses it to estimate lateral camera-frame velocities:

$c_i$7

Longitudinal velocity is estimated from disparity change:

$c_i$8

The resulting per-obstacle variables include $c_i$9, $s_i$0, mask area and shape descriptors, disparity statistics, reconstructed 3D position $s_i$1, velocities $s_i$2, time-to-collision, lateral offset, size or extent, and confidence scores.

A central feature of this formulation is the use of instance masks to restrict disparity and flow aggregation to obstacle pixels rather than whole-image neighborhoods. The robust statistics recommended for these obstacle-centric regions are medians and interquartile ranges, with invalid disparity or flow pixels excluded. This is significant because disparity constrains range along the viewing direction while optical flow provides lateral and longitudinal motion cues from temporal image displacements. The combination yields full 3D motion estimates and threat-oriented interpretation using passive vision alone.

3. Cooperative sensing and geometry optimization

In the multi-UAV formulation, VFEO operates in an inertial East-North-Up coordinate system and estimates the obstacle state

$s_i$3

where $s_i$4 and $s_i$5 (Wang et al., 29 Aug 2025). The obstacle dynamics are modeled as a discrete Markov process:

$s_i$6

with

$s_i$7

Sensing is performed using 5G OFDM DM-RS signals. For UAV $s_i$8, the per-UAV channel observation is

$s_i$9

where $B_i$0 is two-way delay and $B_i$1 is Doppler. Range and radial velocity are then estimated via IFFT and FFT on the nonuniform DM-RS grids. The likelihood model produces individual UAV CRLBs for delay and Doppler, which are converted to range and radial velocity lower bounds:

$B_i$2

The fusion stage uses range and radial-velocity differences relative to the MUAV:

$B_i$3

$B_i$4

The Fisher Information Matrix is

$B_i$5

with $B_i$6 assembled from the per-UAV CRLB-derived variances. The resulting joint lower bound is

$B_i$7

and the position and velocity bounds are summarized by

$B_i$8

VFEO is triggered when the predicted position bound exceeds a threshold, specifically when $B_i$9. The algorithm then solves a constrained nonlinear program whose decision variables are the UAV positions on a formation circle of radius $M_i$0 around the virtual leader. The objective is to minimize $M_i$1 subject to circle adherence, altitude matching with the virtual leader, maximum-speed limits, inter-UAV separation, and obstacle safety distance. The paper converts this to an exterior-penalty surrogate and solves it using steepest descent:

$M_i$2

The significance of this construction is geometric. The method does not attempt to improve the estimator solely by refining signal processing; it also improves the conditioning of the Fisher information through variable formation in the spatial arrangement of sensors. In the paper’s terms, VFEO increases geometric diversity and reduces dilution of precision when a uniform formation yields poor sensitivity.

4. Fusion, tracking, and hierarchical control

The passive-vision blueprint extends beyond feature extraction to temporal fusion and tracking (Rateke et al., 2020). Its recommended enhancement is a constant-velocity Kalman filter per tracked obstacle with state

$M_i$3

measurement vector

$M_i$4

and standard predict-update recursions. Measurement uncertainty is propagated from disparity spread, flow variability, and disparity-rate variability; measurement noise is then scaled by disparity confidence, flow confidence, and detection confidence. Egomotion compensation is implemented by estimating background flow outside all masks and subtracting it from each obstacle’s median flow. Time-to-collision is defined as

$M_i$5

for approaching objects with $M_i$6.

The multi-UAV formulation couples sensing with path-following and avoidance through a hierarchical fusion architecture (Wang et al., 29 Aug 2025). Obstacle state estimation is produced at the MUAV using a two-step weighted least squares (TWLS) estimator, while formation-following is handled by an adaptive-weight DDPG controller with reward

$M_i$7

where $M_i$8 and $M_i$9. Obstacle avoidance is triggered when $D(u,v)=x_l-x_r,$0 and uses

$D(u,v)=x_l-x_r,$1

Conflict-free fusion is then achieved by the null-space-based hierarchical strategy

$D(u,v)=x_l-x_r,$2

where $D(u,v)=x_l-x_r,$3 is the VFEO sensing-induced reconfiguration velocity and $D(u,v)=x_l-x_r,$4 is the path-following velocity.

These two formulations use different estimation architectures, but both rely on the same structural principle: raw observations are converted into obstacle-specific state variables, uncertainties are modeled explicitly, and downstream control or threat assessment consumes the fused state rather than the original measurements. This suggests a unifying interpretation of VFEO as an estimation-and-control interface rather than only a sensing module.

5. Empirical evaluation

For road-obstacle perception, the evaluation uses KITTI and CaRINA, with 100 frames in 20 sequences, 5 frames each, and 415 obstacles manually annotated for depth and motion labels (Rateke et al., 2020). The reported accuracies are 81.75% for depth labeling, 89.51% for x-axis direction, 83.57% for y-axis direction, and 80.96% for movement intensity. The depth labels are very close, close, far, and very far; the x-axis direction labels are left-to-right, right-to-left, and stable; the y-axis direction labels are approaching, moving-away, and stable; and the movement-intensity labels are stopped, slow, average, fast, and very fast. The confusion matrices are described as dominated by near-neighbor errors, and qualitative examples include a truck crossing right-to-left, pedestrians crossing with similar but unsynchronized motions, and scenes with stationary ego vehicle and crossing traffic.

For cooperative sensing, the simulation setup uses $D(u,v)=x_l-x_r,$5 UAVs, with one MUAV and four AUAVs, formation radius $D(u,v)=x_l-x_r,$6, inter-UAV safety distance $D(u,v)=x_l-x_r,$7, obstacle safety distance $D(u,v)=x_l-x_r,$8, and CRLB threshold $D(u,v)=x_l-x_r,$9 (Wang et al., 29 Aug 2025). The ISAC configuration uses 5G NR DM-RS at $d_t(u,v)$0, $d_t(u,v)$1, $d_t(u,v)$2, $d_t(u,v)$3, $d_t(u,v)$4, with DM-RS on $d_t(u,v)$5 subcarriers and $d_t(u,v)$6 OFDM symbols, under SNR $d_t(u,v)$7. The AWPF controller converges in approximately 160 episodes, achieves average path-following error below $d_t(u,v)$8, and improves following accuracy by 21%–124% relative to the listed baselines. For sensing, fixed-formation estimation errors of $d_t(u,v)$9 and $(u_t(u,v),v_t(u,v))$0 at 276 s and 278 s are reduced to $(u_t(u,v),v_t(u,v))$1 and $(u_t(u,v),v_t(u,v))$2 after VFEO reconfiguration. The paper further reports that avoidance is triggered at 2 s for one obstacle, lasts 15 s, remains collision-free, and that the path-following error returns to baseline within 4 s post-avoidance.

Taken together, these results show two distinct performance profiles. In passive road scenes, the emphasis is on categorical correctness of depth and motion labels extracted from passive imagery. In cooperative UAV sensing, the emphasis is on CRLB reduction and meter-to-decimeter improvements in continuous position estimation through formation reconfiguration.

6. Limitations, misconceptions, and prospective extensions

The passive-vision blueprint has several stated failure modes (Rateke et al., 2020). Low-texture or specular surfaces can make disparity and flow unreliable; motion blur and sudden rotations can make background-flow compensation insufficient; occlusions and crowded scenes can cause mask overlap and disparity bleeding across borders; slanted or thin structures can produce stereo bias and outliers; calibration and baseline errors directly affect $(u_t(u,v),v_t(u,v))$3 and therefore propagate to TTC; and small or near objects can yield partial masks with too few valid pixels. The recommended responses are increased uncertainty, downweighting, validity-ratio thresholds, robust medians, instance-mask erosion or dilation, and temporal smoothing. Comparative context is also explicit: monocular-only methods lack absolute depth, whereas stereo disparity provides metric $(u_t(u,v),v_t(u,v))$4; active sensors such as LiDAR and radar provide robust depth and velocity, especially in low texture, but introduce higher cost and complexity.

The multi-UAV algorithm likewise depends on explicit assumptions (Wang et al., 29 Aug 2025). These include line-of-sight availability, strict inter-UAV clock synchronization, Gaussian measurement noise, independence across UAVs in measurement noise, accurate handling of nonuniform DM-RS indices, Markov obstacle dynamics, a virtual leader model, and a formation constrained to a circle at fixed altitude. The paper notes that multipath and non-line-of-sight can bias range and Doppler estimates, communication latency or synchronization errors can affect difference measurements, model mismatch can degrade fusion performance, and the steepest-descent penalty solver may converge to local minima depending on the penalty weight, step sizes, and stopping criterion. A further limitation is that the fixed formation-circle constraint may limit attainable optimal geometries in highly asymmetric scenarios.

A common misconception is that VFEO denotes a single standardized algorithm independent of sensing modality. The record here indicates something narrower and more technical: the term explicitly names a CRLB-driven variable-formation method for multi-UAV cooperative sensing, while a passive-vision formulation can be interpreted as a VFEO-style blueprint because it uses object-centric variable formation and fusion. Another misconception is that passive sensing can only provide weak qualitative cues. The reported road-scene results show that passive stereo and optical flow, when aggregated within instance masks, can support class, position, depth, and motion extraction with nontrivial accuracy. Conversely, it would also be incorrect to assume that geometry optimization alone solves obstacle sensing; the UAV formulation remains sensitive to propagation conditions, synchronization, and model assumptions.

The prospective extensions stated in the source material are consistent with these limitations. For passive vision, they include visual odometry or IMU for egomotion compensation, learning-based stereo and flow such as PWC-Net or RAFT in low-texture scenes, multi-hypothesis tracking for split or merge occlusions, and adaptive fusion through scene-dependent adjustment of $(u_t(u,v),v_t(u,v))$5 and $(u_t(u,v),v_t(u,v))$6. For cooperative sensing, the listed directions include distributed VFEO optimization, direct A-/D-/E-optimal active sensing criteria, adaptive waveform design to optimize the Fisher Information Matrix, and robust fusion under NLOS or multipath using EKF or UKF with augmented channel models. These directions preserve the central VFEO principle: obstacle estimation quality is governed jointly by the variables formed from measurements and by the geometry, uncertainty model, and fusion mechanism through which those variables are used.