mmE-Loc: Precise Drone Landing via Event Camera & Radar

• The paper introduces mmE-Loc, a novel sensor fusion system combining an event camera and mmWave radar to achieve ultra-high-frequency localization for precise drone landing.
• It addresses sensor-rate mismatch by replacing conventional frame cameras with asynchronous event cameras, ensuring ≥150 Hz pose updates and <10 ms latency for effective control.
• Extensive experiments demonstrate mmE-Loc attains sub-decimeter accuracy and 50–60% error reduction over baselines under various indoor and outdoor conditions.

Searching arXiv for the cited paper and closely related baseline work. arxiv_search(query="mmE-Loc Facilitating Accurate Drone Landing with Ultra-High-Frequency Localization", max_results=5, sort_by="relevance") Searching for event-camera drone localization and radar-camera fusion baselines mentioned in the source material. arxiv_search(query="Falanga event camera drone localization arXiv", max_results=10, sort_by="relevance") arxiv_search(query="milliEye frame camera radar IoTDI 2021 arXiv", max_results=10, sort_by="relevance") arxiv_search(query="radar only tracker Zhao ICRA 2021 drone localization arXiv", max_results=10, sort_by="relevance") arxiv_search(query="event camera radar fusion localization arXiv", max_results=10, sort_by="relevance") mmE-Loc is a ground localization system for precise drone landing that combines an event camera with an mmWave FMCW radar in order to satisfy the simultaneous requirements of high spatial precision, high update rate, and low end-to-end latency during descent guidance. It is introduced in "mmE-Loc: Facilitating Accurate Drone Landing with Ultra-High-Frequency Localization" (Wang et al., 13 Jul 2025) as a response to the sampling-frequency mismatch between conventional frame cameras and radar: frame cameras operate at $\leq 50$ Hz with exposure time $>20$ ms, whereas mmWave radar can operate at up to $200$ Hz or more. The central design choice is to replace the frame camera with a bio-inspired event camera that reports per-pixel brightness changes asynchronously with millisecond latency and effectively samples at kHz rates, then to fuse event-stream image information and radar depth measurements through two modules, Consistency-instructed Collaborative Tracking and Graph-informed Adaptive Joint Optimization, for accurate landing-site localization.

1. Problem setting and operating requirements

mmE-Loc is formulated for real-time localization of descending drones by a ground platform that must guide the vehicle to a designated landing spot. The stated landing requirement is stringent: the pad should place drones within a few centimeters, specifically $\leq 10$ cm, of a charging port to avoid damage or safety risks (Wang et al., 13 Jul 2025). The control-side requirement is similarly explicit: drone flight controllers typically require pose updates $\geq 150$ Hz to maintain stable descent trajectories under closed-loop control, while sensing and computation should introduce only a few milliseconds of latency, ideally $<10$ ms.

The system addresses a specific systems bottleneck. Conventional frame cameras are too slow relative to mmWave radar, so naïve fusion forces the overall update rate down to the camera rate. mmE-Loc therefore targets both throughput and accuracy simultaneously rather than treating sensor fusion only as a perception problem. A common misconception is that radar-camera fusion is limited primarily by estimator design; the reported formulation instead identifies sensor-rate harmonization as a first-order systems constraint.

This suggests that mmE-Loc should be understood not merely as a multimodal localization algorithm, but as a co-design of sensing stack and inference stack. The event camera is introduced precisely because its asynchronous measurement regime better matches the temporal granularity of radar updates.

2. Sensor configuration and measurement models

The hardware configuration consists of a Prophesee EVK4 HD event camera in frame $E$ and a TI IWR1843 FMCW radar in frame $R$, with a known rigid transform $T_{ER}$ between them (Wang et al., 13 Jul 2025). The event camera has resolution $1280 \times 720$ pixels and a $>20$0 field of view, and outputs events of the form $>20$1 with polarity $>20$2. The radar operates at $>20$3--$>20$4 GHz and uses $>20$5 Tx and $>20$6 Rx antennas arranged in two orthogonal linear arrays. The coordinate frames $>20$7 and $>20$8 are static, whereas the object or drone frames $>20$9 and $200$0 have unknown translations $200$1 or $200$2 at time index $200$3.

For radar ranging, the transmitted FMCW chirp is

$200$4

and the received echo from a target at distance $200$5 is modeled as a delayed, attenuated copy,

$200$6

After mixing and low-pass filtering, the intermediate-frequency tone is

$200$7

with beat frequency $200$8, yielding

$200$9

Angular estimation uses phase differences $\leq 10$0 between adjacent receive elements spaced by $\leq 10$1:

$\leq 10$2

A unit direction vector is then formed as

$\leq 10$3

Each radar point in radar coordinates is

$\leq 10$4

and in camera coordinates

$\leq 10$5

A simple constant-velocity Kalman filter tracks $\leq 10$6 from successive $\leq 10$7.

For the event camera, each pixel $\leq 10$8 generates an event

$\leq 10$9

whenever its log intensity changes by a threshold $\geq 150$0. A 3D point $\geq 150$1 projects to image coordinates via the pinhole model

$\geq 150$2

using intrinsics $\geq 150$3. Conversely, a bounding-box center $\geq 150$4 can be back-projected along a camera ray if depth is known.

These models establish the modal complementarity exploited by mmE-Loc: radar contributes explicit depth and direction, while the event camera contributes high-rate image-plane information with millisecond latency.

3. Consistency-instructed Collaborative Tracking

Consistency-instructed Collaborative Tracking, or CCT, is the front-end module that jointly processes noisy high-rate radar and event outputs to identify the landing drone and extract basic 3D measurements (Wang et al., 13 Jul 2025). Its logic is based on two cues: cross-modal temporal consistency and drone-specific physical signatures.

The temporal-consistency rule is simple but restrictive. Radar points and event-based detections produced by the same object should coincide in time within a few milliseconds and should lie along the same camera ray. The system projects each radar point $\geq 150$5 into the event camera and retains only radar returns that fall within current event bounding boxes. Conversely, event clusters without support from any radar return are discarded. This makes CCT a mutual-gating mechanism rather than a unilateral proposal stage.

CCT also uses the physical structure of a descending multicopter. The drone is described as the only object producing both high-frequency micro-motion bursts from fast propeller rotation and axial-symmetry patterns in the spatial layout of events, with $\geq 150$6- or $\geq 150$7-propeller-arm structure. Over a short window $\geq 150$8, events are binned into a $\geq 150$9D histogram $<10$0, for example with $<10$1-pixel bins. Candidate bins satisfy

$<10$2

capturing dense bipolar event activity generated by true propellers. Connected-component clusters of these hot bins are then ellipse-fitted, and only clusters exhibiting one or more axes of symmetry consistent with $<10$3- or $<10$4-fold propeller layouts are accepted as drone.

The CCT output is a preliminary estimate of the drone translation $<10$5 in the camera frame, together with a filtered event cluster center $<10$6 and the corresponding radar tuple $<10$7 for downstream fusion. The reported ablation results indicate that these physically informed filters are not incidental engineering details: CCT precision and recall are both above $<10$8 on both modalities, and the Radar + CCT configuration achieves error $<10$9 m with latency $E$0 ms.

A plausible implication is that CCT functions as a domain-specific association prior. Instead of learning object identity from generic appearance, it encodes periodicity and symmetry that are specific to multirotor descent, which is particularly useful in cluttered landing environments.

4. Graph-informed Adaptive Joint Optimization

Graph-informed Adaptive Joint Optimization, or GAJO, is the back-end fusion and smoothing module. It estimates the set of drone poses

$E$1

from measurements

$E$2

by maximizing the posterior

$E$3

which yields a non-linear least-squares problem (Wang et al., 13 Jul 2025).

The factorization includes three types of terms. A prior factor imposes constant-velocity motion. An event likelihood factor penalizes reprojection error between $E$4 and the image observation $E$5. A radar likelihood factor incorporates consistency with measured range $E$6, angle $E$7, and frame-to-frame motion $E$8. In operational terms, the graph combines geometric agreement with dynamic regularization.

Optimization proceeds in two stages. Inter-SAE Tracking performs an instantaneous update in which only the current pose $E$9 is optimized while the history is fixed; this runs at every event/radar update and is intended for low latency. Local Location Optimization performs batch reoptimization over a sliding window of the last $R$0 poses,

$R$1

every few frames or when the IMU-predicted pose diverges by more than $R$2.

GAJO is explicitly motion-aware. The drone’s onboard IMU acceleration is used to predict $R$3, and the deviation $R$4 triggers the batch-window update. The window size $R$5 is adjusted dynamically to balance drift correction against computation. Incremental inference uses an iSAM-style strategy, specifically the ISAM2 approach, so that only affected parts of the QR-factorized system are updated instead of relinearizing the entire window at each step. The stated worst-case per-step complexity is $R$6 for window size $R$7, while typical incremental steps cost $R$8.

The ablation study positions GAJO as more than a conventional filter replacement. Radar + GAJO reaches error $R$9 m with latency $T_{ER}$0 ms, and the full system improves error by $T_{ER}$1--$T_{ER}$2 relative to EKF or graph-only alternatives. The design therefore treats low latency and temporal consistency as joint optimization targets rather than separate concerns.

5. Empirical performance and robustness

The experimental campaign spans indoor and outdoor landing scenarios. Indoor experiments use an $T_{ER}$3 capture volume with Vicon motion capture at $T_{ER}$4 mm accuracy. Outdoor experiments use RTK-GPS reference at $T_{ER}$5 cm accuracy. The tested drones are DJI Mini 3 Pro, MAVIC 2, and M30T, covering varying size and propeller count. Conditions include bright and dim indoor lighting, moving balls and partial occlusions as background dynamics, slow descent and spiral trajectories at $T_{ER}$6 m, and velocities up to $T_{ER}$7 m/s. The reported dataset exceeds $T_{ER}$8 h and $T_{ER}$9 GB raw (Wang et al., 13 Jul 2025).

The evaluation compares mmE-Loc to four baselines: single-chip radar only + tracker [Zhao et al. ICRA’21], mono-event 3D localization with known geometry [Falanga et Sci. Robots ’20], stereo-event depth estimation adapted from Falanga et al., and a deep-fusion CNN using frame camera + radar [milliEye IoTDI’21]. The reported results indicate that mmE-Loc outperforms these baselines in both accuracy and latency.

Setting	mmE-Loc	Baselines / note
Indoor 3D error, 90th percentile	$1280 \times 720$0 m	$1280 \times 720$1 m, $1280 \times 720$2 m, $1280 \times 720$3 m, $1280 \times 720$4 m
Indoor 3D error, mean	$1280 \times 720$5 m	reported with 90th-percentile values above
Outdoor average error (M30T)	$1280 \times 720$6 m	$1280 \times 720$7 m--$1280 \times 720$8 m
End-to-end latency, indoor mean	$1280 \times 720$9 ms	$>20$00--$>20$01 ms
End-to-end latency, outdoor mean	$>20$02 ms	$>20$03--$>20$04 ms

Robustness results are reported across multiple nuisance factors. Across different drones, mmE-Loc maintains error $>20$05 m, whereas the baselines degrade heavily when geometry changes. Under dim illumination, mmE-Loc remains at $>20$06 m, while event-only methods degrade. Under heavy background clutter, average error is $>20$07 m, whereas radar-only and event-only methods drift to $>20$08 m. By distance, the average error is $>20$09 m for $>20$10 m, $>20$11 m for $>20$12--$>20$13 m, and $>20$14 m for $>20$15 m. Under partial occlusion of $>20$16--$>20$17, error remains $>20$18 m. By velocity band, the average error is $>20$19 m for $>20$20 m/s, $>20$21 m for $>20$22--$>20$23 m/s, and $>20$24 m for $>20$25--$>20$26 m/s.

Ablation results isolate the contribution of modality fusion and module design. Combining modalities reduces error by $>20$27--$>20$28 relative to radar-only or event-only systems. Radar + CCT yields error $>20$29 m at $>20$30 ms latency; Radar + GAJO yields $>20$31 m at $>20$32 ms; full mmE-Loc yields $>20$33 m at $>20$34 ms. Resource usage is modest: CPU utilization $>20$35 and memory $>20$36 MB on a $>20$37-core i7 with GTX1070 GPU.

These figures support a specific interpretation of the method’s contribution. The best operating point is not obtained by maximizing back-end sophistication alone or by maximizing front-end selectivity alone; it emerges from the interaction between a physically constrained front-end and an adaptive incremental graph back-end.

6. Scope, limitations, and relation to adjacent approaches

The reported system is explicitly limited to a single-pad scenario with one landing drone at a time (Wang et al., 13 Jul 2025). Multi-pad or simultaneous-landing extensions require track ID management. It is also a ground-station deployment rather than an onboard localization system; an onboard analogue would require motion compensation in the event stream and denser radar returns. The experiments evaluate mmE-Loc as a standalone system rather than in tight coupling with GPS/RTK or visual markers, although such coupling is identified as a possible route to increased reliability in urban canyons.

Environmental boundaries are also clearly stated. Heavy rain, fog, and RF interference remain unevaluated. Generalization to other fast-rotating platforms, including VTOL multicopters, would require parameter tuning for propeller frequencies and symmetry. These are not incidental caveats: several parts of CCT rely directly on periodicity and axial symmetry, so deployment beyond the tested operating envelope is not guaranteed by the present results.

Within the broader multimodal localization landscape, mmE-Loc occupies a specific position. It is not a radar-only tracker, not an event-only estimator, and not a frame-camera-plus-radar deep-fusion pipeline. Its stated contribution is to use event cameras and mmWave radars at matched high rates in order to break the camera-radar throughput bottleneck, while also using physical priors and adaptive graph optimization to maintain millisecond-scale latency on commodity hardware. The reported evidence therefore supports viewing mmE-Loc as an integrated sensing-and-inference architecture for precision landing rather than as an isolated fusion module.

In summary, mmE-Loc couples ultra-high-rate event imaging with mmWave depth sensing, uses CCT to extract drone-consistent measurements through temporal consistency and propeller-aware structure, and uses GAJO to fuse these measurements into a smooth 3D trajectory at low latency. The resulting system is reported to satisfy the practical descent-guidance regime of $>20$38 Hz updates, few-millisecond latency, and sub-decimeter average localization error in the tested landing scenarios (Wang et al., 13 Jul 2025).