Proximity Monitoring and Warning (PMW)

• PMW is a system that continuously monitors the spatial and temporal distances between entities using sensors like RFID, UWB, and computer vision to prevent collisions.
• It employs graded warning strategies—including visual, auditory, and vibrotactile alerts—with risk metrics based on time-to-collision and stopping distance criteria.
• Evolving from early RFID and GPS methods to advanced deep learning and sensor fusion, PMW enhances safety across construction sites, roadwork zones, and human-centered robotics.

Searching arXiv for the cited PMW papers and related surveys to ground the article in current literature. Proximity Monitoring and Warning (PMW) denotes a class of systems that continuously monitor proximity between entities and issue warnings or control actions when that proximity becomes hazardous. In construction, PMW refers to systems and methods that continuously monitor the proximity between pairs of construction entities and issue warnings in order to prevent struck-by accidents and enhance onsite safety (Ding et al., 17 Jul 2025). In human-centered robotics, proximity perception detects objects without physical contact, typically at short range up to about $50\,\text{cm}$, with a “pre-touch” sub-range below about $10\,\text{cm}$, bridging the gap between long-range vision and contact sensing (Navarro et al., 2021). Across the literature, proximity is represented as geometric distance, time to a collision zone, time-to-collision, protective separation distance, or context-aware exposure, and PMW implementations range from advisory alarms to speed modulation, protective stop, and autonomous braking (Gelbal et al., 2023, Yu et al., 2024).

1. Scope and evolution

PMW has developed as a cross-domain safety paradigm spanning construction, road traffic, work zones, collaborative robotics, industrial sensing, healthcare simulations, and wearable safety systems. In construction, the surveyed literature covers 97 PMW articles published between 2010 and 2024 and treats workers, construction equipment and vehicles, materials and loads, and temporary or permanent structures and utilities as the principal “construction entities” whose relative states can create collision risks (Ding et al., 17 Jul 2025). In highway work zones, a parallel review of 63 publications frames the main safety functions as localization, vehicle intrusion detection, and worker–equipment proximity detection and warning, emphasizing that work zones are temporary, spatially constrained, noisy, and often power-limited (Demeke et al., 5 Mar 2025).

The technical trajectory is staged. From 2010 to 2016, PMW work was dominated by RFID, GPS, laser scanning, and blind-spot analysis. From 2017 to 2019, UWB and BLE increasingly replaced RFID for more precise positioning, while UAV-based monitoring and early AI methods appeared. From 2020 onward, computer vision and deep learning became dominant, with YOLO-family detectors, trajectory prediction, and vibrotactile warning emerging as characteristic themes (Ding et al., 17 Jul 2025). A related survey in human-centered robotics places proximity sensing within two principal scenarios—robot exteriors for safety and interaction, and grippers or hands for grasping and exploration—and formalizes proximity sensing as a short-range, low-latency, tightly control-coupled layer (Navarro et al., 2021).

The application scope has also broadened. Road-safety implementations monitor vulnerable road users and oncoming vehicles; robotic systems monitor humans, obstacles, and workspace occupancy; occupational systems track workers against hazard fields; healthcare-oriented systems monitor face-to-face contacts and room-level presence; and dementia-oriented systems treat proximity as a context-aware relation between an elder and familiar areas or hazardous scenes (Gelbal et al., 2023, Mitchell et al., 2022, Ozella et al., 2018, Deng et al., 2024).

2. Hazard representation and decision logic

The central PMW problem is the conversion of spatial and temporal state estimates into hazard levels. Construction PMW literature distinguishes static hazard areas from dynamic hazard areas. Static areas include concentric circles or fans, concentric rectangles, ellipses, irregular polygons, and polygonal virtual fences, while dynamic areas include dynamic workspaces, sweep-volume-based hazard regions, and egg-shaped zones whose geometry changes with equipment pose, speed, braking time, and worker response assumptions (Ding et al., 17 Jul 2025).

A canonical temporal-proximity formulation appears in smartphone-based vehicle-to-pedestrian PMW. There, a collision zone is a rectangular region where straight-line pedestrian and vehicle paths intersect, and proximity is expressed by Time-to-Zone (TTZ): $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$ where $d_y$ and $d_p$ are distances to the collision zone along the respective paths and $v_y$, $v_p$ are speeds. Collision risk is flagged when the vehicle and pedestrian are predicted to reach the zone within a safety window: $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$ with $t_s = 4\,\text{s}$ in the reported implementation (Gelbal et al., 2023).

Roadway PMW often uses stopping-based criteria. For truck-mounted attenuators, the real-time warning threshold is derived from AASHTO stopping sight distance: $\text{SSD} = 1.47 V t + \frac{1.075 V^2}{a}$ where $10\,\text{cm}$0 is vehicle speed in mph, $10\,\text{cm}$1 is brake reaction time, and $10\,\text{cm}$2 is deceleration rate. A warning is issued when measured distance falls below $10\,\text{cm}$3 (Yu et al., 2024). At roundabouts, infrastructure-based PMW uses predicted time of potential collision, time headway to the conflict point, and the Crash Potential Index (CPI), which integrates the probability that the deceleration required to avoid a crash exceeds the available deceleration rate (Zhang et al., 2023).

Human-centered robotics uses a safety-standard formulation. For Speed and Separation Monitoring, the protective separation distance is written as

$10\,\text{cm}$4

where $10\,\text{cm}$5 and $10\,\text{cm}$6 are human and robot speeds, $10\,\text{cm}$7 is reaction time, $10\,\text{cm}$8 is stopping time, $10\,\text{cm}$9 is stopping distance, and $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$0, $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$1, $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$2 capture intrusion and uncertainty terms (Navarro et al., 2021). Construction PMW extends thresholding with continuous risk quantities such as warning indices, fuzzy risk degrees, spatio-temporal network risk, proximity hazard indicators, and hazard exposure integrals over time (Ding et al., 17 Jul 2025). This suggests that PMW is best understood not as a single metric family but as a layered decision framework that selects among distance, time, stopping feasibility, visibility, and exposure according to domain constraints.

3. Sensing and perception modalities

PMW systems are sensor-diverse. In construction and work zones, radio and localization modalities include RFID, BLE, UWB, CSS, GNSS/GPS, and IMUs. UWB is reported at about $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$3 accuracy in construction contexts, CSS-based RTLS achieves about $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$4 accuracy with $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$5 latency, BLE offers sub-meter accuracy under good conditions, and GPS-aided INS reaches about $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$6 localization accuracy while reducing false alarms relative to GPS alone (Ding et al., 17 Jul 2025). In highway work zones, passive RFID ranges from a few centimeters up to about $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$7, active RFID up to about $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$8, BLE typically $$TTZ_y = \frac{d_y}{v_y}, \qquad TTZ_p = \frac{d_p}{v_p}$$9–$d_y$0, and UWB offers $d_y$1–$d_y$2 or centimeter-level positioning with sub-millisecond latency, albeit at higher cost and power draw (Demeke et al., 5 Mar 2025).

A distinct communication-centric PMW architecture appears in smartphone-based vehicle–pedestrian safety. Two Android applications exchange SAE J2735 Personal Safety Messages over Bluetooth 5.0 BLE extended advertisements using CODED PHY, with the pedestrian phone acting as transmitter and the vehicle phone as receiver and PMW engine. The implementation uses only GPS-derived latitude, longitude, speed, and heading, performs all processing on-device, and demonstrates early warning in no-line-of-sight conditions behind a building (Gelbal et al., 2023). WiFi CSI-based indoor PMW uses a different logic: adjacent-subcarrier correlation as a proximity feature, gait monitoring from the autocorrelation of CSI power, and a four-state finite-state machine. Across four environments, that system reports an overall detection rate of $d_y$3, a false alarm rate of $d_y$4, and an average delay of $d_y$5 (Hu et al., 2024).

Vision-based PMW spans monocular, stereo, panoptic, and infrastructure sensing. In construction, a monocular 2D camera system couples PGD-based monocular 3D detection with rule-based proximity categorization into Dangerous, Potentially Dangerous, Concerned, and Safe; the trained model achieves $d_y$6 loose AP within $d_y$7, and the implemented PMW system achieves an $d_y$8 of roughly $d_y$9 within $d_p$0 under specified settings (Ding et al., 2023). For mobile work zones, a ROS-distributed system uses YOLOP for object detection, drivable-area segmentation, and lane-line segmentation, fuses RGB and depth images, computes distance from the central third of the vehicle bounding box, estimates relative speed from successive distance samples, and drives LED warnings on a follower truck surrogate (Yu et al., 2024). Infrastructure-based roundabout PMW assumes roadside cameras, LiDARs, or radars plus V2X and evaluates warning timing in a human-in-the-loop co-simulation environment (Zhang et al., 2023).

Radar, acoustics, and near-field surface sensing extend PMW into conditions where optics are weak. Millimeter-wave FMCW radar has been used as a foresight sensor that distinguishes humans from metallic infrastructure, detects a human-carried $d_p$1 copper sheet through a partition wall, and supports through-wall awareness in shared workspaces and beyond doorways (Mitchell et al., 2022). MIRO uses multiple overlapping $d_p$2–$d_p$3 FMCW radars plus PM sensors; after GAN-based view adaptation and cross-radar matching, it achieves a re-identification $d_p$4-score of $d_p$5 and mean SSIM of $d_p$6 for view adaptation, enabling identity-aware worker tracking in dusty industrial settings (Halder et al., 8 Mar 2026). AuraSense uses a Leaky Surface Wave on a robot manipulator and reports $d_p$7 true positive proximity detection for static obstacles, $d_p$8 for mobile obstacles, and a true negative rate over $d_p$9 with only one pair of piezoelectric transducers (Fan et al., 2021). Roadside acoustic PMW classifies heavy vehicles, light vehicles at low or high speed, and no-vehicle sound using MFCC, LPC, FFT-derived features, and an MLP, reaching event-level accuracies from $v_y$0 to $v_y$1 (Khalili et al., 2018). AirTouch combines an IR camera with an AprilTag-based wearable marker and an airflow barrier, estimating marker pose at $v_y$2 per frame and using the human–robot distance to trigger tactile warning via air pressure (Rakhmatulin et al., 2023).

4. Warning semantics, interfaces, and actuation

PMW warning logic is usually graded rather than binary. In the smartphone vehicle–pedestrian implementation, the vehicle phone has four states—no warning, yellow warning, orange warning, and red warning. Yellow and orange depend primarily on decreasing vehicle TTZ, while red is tied to braking feasibility through the minimum required deceleration

$v_y$3

and is triggered when $v_y$4 exceeds the maximum braking deceleration assumed by the system (Gelbal et al., 2023). In a related automotive prototype for pedestrians and cyclists, an early warning is defined as a warning issued at least $v_y$5 before an emergency warning; the algorithm searches for path intersections within a $v_y$6 horizon and up to $v_y$7 ahead of the vehicle’s route (Wolf, 2021).

Warning delivery modalities include visual, auditory, vibrotactile, haptic-airflow, and direct control actuation. Construction PMW studies use in-cab displays, LED strips on helmets or equipment, AR glasses, speakers, buzzers, and vibrotactile motors embedded in belts, vests, helmets, or wristbands (Ding et al., 17 Jul 2025). Visual warnings alone can improve safety: in a roundabout simulator with 36 participants, advanced infrastructure-based warnings reduced crash potential, and $v_y$8 warnings yielded smoother deceleration than $v_y$9 warnings (Zhang et al., 2023). Auditory warnings remain common but are susceptible to noise and habituation; vibrotactile warnings become attractive precisely because construction sites are acoustically hostile, and visual plus vibrotactile warnings have been reported to outperform auditory alarms in response time (Ding et al., 17 Jul 2025).

The warning channel can itself embody a spatial model. AirTouch activates airflow when the hand–TCP distance falls below the haptic activation distance of $v_p$0, with $v_p$1 as a dangerous-proximity reference. Participants could discriminate these two radii using airflow alone, and in an inattentive human–robot interaction scenario the average distance to the robot in the dangerous zone increased from $v_p$2 to $v_p$3 when air feedback was enabled (Rakhmatulin et al., 2023). Other systems actuate warnings more directly: the TMA prototype writes a Boolean warning state to /led/status, and the elderly anti-loss helmet escalates from SMS to pop-up to direct phone call according to four danger levels, Safe, Low, Medium, and High (Yu et al., 2024, Deng et al., 2024).

Alarm suppression and timing are critical because redundant warnings create habituation and alarm fatigue. Construction PMW therefore includes field-of-vision-aware suppression, direction-aware warning logic, TTC gating, and continuous-risk thresholds rather than simple distance-only triggers (Ding et al., 17 Jul 2025). The indoor WiFi PMW system explicitly uses a finite-state machine to preserve the Near state during subtle motion and avoid flickering warnings when motion energy drops (Hu et al., 2024). A plausible implication is that the warning problem is no longer merely one of detecting proximity; it is increasingly one of deciding which warnings should be rendered, when, and through which channel, under bounded nuisance-alarm rates.

5. Application domains and operational forms

Road-traffic PMW focuses on vulnerable road users, intersections, and work zones. Smartphone P2V systems use phones as distributed V2X nodes and explicitly address reduced-visibility and non-line-of-sight crosswalk scenarios (Gelbal et al., 2023). Early-warning automotive prototypes use onboard ADAS cameras, EKF-based vehicle-state estimation, cloud-derived probable paths, and directional audio prompts for pedestrians and cyclists (Wolf, 2021). Infrastructure-based PMW at roundabouts moves sensing and prediction to the roadside, then broadcasts warnings via V2X to the ego vehicle (Zhang et al., 2023). Mobile work-zone PMW adapts similar logic to rear-end threats against attenuator trucks, with lane-aware perception, SSD-based warning thresholds, and additional LED actuation (Yu et al., 2024). Highway work-zone reviews further organize sensor use into localization backbones, intrusion detection, and worker–equipment proximity warning, emphasizing that BLE-based systems currently offer a practical balance of cost, power, and performance for small and medium-size work zones, while UWB offers the strongest technical performance where centimeter-level accuracy is required (Demeke et al., 5 Mar 2025).

Construction PMW is broader in entity coverage and richer in hazard modeling. The literature addresses worker–equipment, equipment–equipment, equipment–structure, crane–load, crane–crane, equipment–power-line, and excavator–underground-utility scenarios (Ding et al., 17 Jul 2025). Perception ranges from RFID and BLE tags through UWB RTLS and GPS-aided INS to vision systems based on YOLO, Faster R-CNN, Mask R-CNN, monocular depth estimation, stereo vision, and point-cloud registration. Hazard levels may be static, dynamic, discrete, or continuous, and warning delivery increasingly includes vibrotactile channels precisely because struck-by hazards often occur under high cognitive and acoustic load (Ding et al., 17 Jul 2025).

Human-centered robotics treats PMW as near-body protection and pre-touch interaction. Exterior robot skins use capacitive, optical, inductive, radar, acoustic, or full-surface sensing to maintain protective separation and enable avoidance or protective stop (Navarro et al., 2021). Through-wall FMCW radar provides “foresight sensing” in shared workspaces and beyond doors (Mitchell et al., 2022). AuraSense creates a no-dead-spot “aura” around a manipulator link, while AirTouch uses airflow as a physically felt warning field around a collaborative robot (Fan et al., 2021, Rakhmatulin et al., 2023). These systems illustrate two distinct PMW strategies: one uses the sensor to estimate distance and then modulate robot motion; the other uses the sensor to estimate distance and directly modulate human perception.

The PMW concept also extends beyond direct collision avoidance. In mass-casualty incident exercises, wearable active RFID tags and room beacons reconstruct face-to-face contacts at $v_p$4 resolution and patient presence in rooms over $v_p$5 windows, enabling contact matrices, flow analysis, and bottleneck detection that could support warning logic for under-served critical patients or overloaded zones (Ozella et al., 2018). In dementia-oriented anti-loss systems, GPS, first-person images, 5G connectivity, and a multimodal scene–location network map spatio-temporal context into Safe, Low, Medium, and High danger states, achieving $v_p$6 accuracy in the best reported model and delivering fully automatic caregiver warnings without requiring user interaction from the elder (Deng et al., 2024). MIRO similarly turns identity-aware worker trajectories into worker-specific particulate exposure estimates, showing that PMW architectures can fuse motion tracking with environmental scalar fields rather than with collision risk alone (Halder et al., 8 Mar 2026).

6. Limitations, evaluation, and future directions

PMW remains constrained by sensing uncertainty, calibration drift, environmental variability, and human factors. GPS and raw BLE are vulnerable to multipath and coarse accuracy; LiDAR and cameras remain line-of-sight dependent; monocular vision retains distance-estimation error; ultrasonic sensing degrades in dust, rain, and noisy outdoor conditions; and mmWave radar still faces clutter, multipath, and angular-coverage limitations (Demeke et al., 5 Mar 2025, Ding et al., 17 Jul 2025, Mitchell et al., 2022). Smartphone P2V PMW inherits GPS noise, straight-line-motion assumptions, BLE congestion, and the need for both pedestrian and driver phones to run the application (Gelbal et al., 2023). Construction monocular PMW is affected by orientation-prediction noise, reduced performance beyond $v_p$7, the synthetic-to-real gap, and the lack of explicit velocity modeling (Ding et al., 2023). WiFi CSI proximity detection assumes a single walking person at a time and suffers on very short paths where gait does not stabilize (Hu et al., 2024). Multi-radar re-identification assumes partial overlap and is not validated in dense crowds (Halder et al., 8 Mar 2026).

Evaluation practices are heterogeneous. Reported metrics include 3D AP, precision, recall, $v_p$8, mean localization error, event-based accuracy, duration accuracy, MSE of relative speed estimates, false alarm rate, SSIM, PSNR, and end-to-end latency. Representative values range from $v_p$9 loose AP within $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$0 and $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$1 within $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$2 for monocular camera PMW, to $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$3 event detection with $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$4 false alarm rate for WiFi PMW, $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$5 re-identification $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$6 for multi-radar industrial tracking, $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$7–$TTZ_p - t_s < TTZ_y < TTZ_p + t_s$8 event-level audio classification for roadside acoustic PMW, and $TTZ_p - t_s < TTZ_y < TTZ_p + t_s$9 precision in one LiDAR intrusion-detection test set (Ding et al., 2023, Hu et al., 2024, Halder et al., 8 Mar 2026, Khalili et al., 2018, Demeke et al., 5 Mar 2025). A recurring conclusion is that the field still lacks standardized PMW-specific KPIs spanning detection probability, miss rate, nuisance alarm rate, localization RMSE, end-to-end latency, and user acceptance.

Future directions are unusually consistent across domains. Construction PMW surveys emphasize end-to-end 3D object detection, real-time 3D reconstruction and updating for dynamic scenes, and multimodal fusion as the primary research frontiers (Ding et al., 17 Jul 2025). Roadside and smartphone road-safety systems explicitly point to Kalman filtering, neural networks, richer message fields, more complex geometry, multi-actor coordination, and deeper ADAS or AV integration (Gelbal et al., 2023). Radar-based robotic PMW calls for full range–Doppler–angle processing, micro-Doppler analysis, sensor fusion, and quantitative safety validation under clutter and multipath (Mitchell et al., 2022). Work-zone systems point toward LiDAR upgrades, better panoptic perception, standardized interfaces, cost-aware modular architectures, and regulatory alignment with V2X and smart-work-zone ecosystems (Yu et al., 2024, Demeke et al., 5 Mar 2025). This suggests that the next phase of PMW will be defined less by any single sensor family than by how well systems fuse heterogeneous perception, uncertainty-aware hazard modeling, and human-centered warning design into certifiable, field-robust safety stacks.