TaFall : Balance-Informed Fall Detection via Passive Thermal Sensing Conference: ; 2026
Chengxiao Li, Xie Zhang, Wei Zhu, Yan Jiang, Chenshu Wu · 2026
Read TaFall : Balance-Informed Fall Detection via Passive Thermal Sensing Conference: ; 2026 by Chengxiao Li, Xie Zhang, Wei Zhu, Yan Jiang, Chenshu Wu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09693.txt.
Introduction
Li and four colleagues introduce TaFall, a balance-informed fall detector using privacy-preserving thermal sensing for private indoor settings. Radio-frequency approaches lean on coarse motion cues. TaFall models falls as balance degradation. Appearance-motion fusion reconstructs pose from low-resolution thermal maps. Balance-aware learning and pose-bridged pretraining add robustness. On over 3,000 instances from 35 participants, detection reaches about 98 percent with under 1 percent false alarms. A 27-day deployment across four homes shows near-zero false alarms. A bathroom pilot confirms moisture and thermal robustness.
Catalog record
| CREATOR | Chengxiao Li, Xie Zhang, Wei Zhu, Yan Jiang, Chenshu Wu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 77710 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.09693
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
TaFall : Balance-Informed Fall Detection via Passive Thermal Sensing Conference: ; 2026 Chengxiao Li, Xie Zhang, Wei Zhu, Yan Jiang, Chenshu Wu Abstract . Falls are a major cause of injury and mortality among older adults, yet most incidents occur in private indoor environments where monitoring must balance effectiveness with privacy. Existing privacy-preserving fall detection approaches, particularly those based on radio frequency sensing, often rely on coarse motion cues, which limits reliability in real-world deployments. We introduce TaFall , a balance-informed fall detection system based on low-cost, privacy-preserving thermal array sensing. The key insight is that TaFall models a fall as a process of balance degradation and detects falls by estimating pose-driven biomechanical balance dynamics. To enable this capability from low-resolution thermal array maps, we propose (i) an appearance-motion fusion model for robust pose reconstruction, (ii) physically grounded balance-aware learning, and (iii) pose-bridged pretraining to improve robustness. TaFall achieves a detection rate of 98.26% with a false alarm rate of 0.65% on our dataset with over 3,000 fall instances from 35 participants across diverse indoor environments. In 27 day deployments across four homes, TaFall attains an ultra-low false alarm rate of 0.00126% and a pilot bathroom study confirms robustness under moisture and thermal interference. Together, these results establish TaFall as a reliable and privacy-preserving approach to fall detection in everyday living environments. 1. Introduction Figure 1. Illustration of thermal array temperature maps and the corresponding body pose–based balance representation during a fall event. The yellow box highlights motion blur in the thermal temperature map induced by rapid movement. In the bottom row, the background color of each frame denotes a distinct balance state—stable balance (SB), loss of balance (LoB), and ground impact stage (GIS)—as detailed in Sec. 3 . Falls constitute a major threat to longevity and independent living among older adults, affecting 28–35% of individuals aged 65 years or above each year ( 38 ) . Among these incidents, 20–30% result in mild to severe injuries ( 30 ) , accounting for approximately 10–15% of all emergency department visits ( 52 ) . Beyond physical harm, falls often trigger a persistent fear of falling, leading 15–55% of older adults to restrict daily activities ( 8 ) , which in turn accelerates functional decline and degrades quality of life. Motivated by these risks, fall detection has attracted increasing research attention in recent years. Existing approaches can be broadly grouped into two categories. The first category is wearable-based systems ( 44 ; 14 ; 34 ; 41 ; 3 ; 5 ) , which can achieve reliable detection performance while preserving privacy. However, their practical adoption is often limited by user adherence issues, including charging requirements, wearing discomfort, and frequent forgetfulness. The second category is non-contact systems that monitor falls without requiring on body devices ( 16 ; 47 ; 49 ; 40 ; 56 ; 58 ) . Among them, camera-based approaches ( 16 ; 47 ; 49 ) can deliver strong accuracy, but raise substantial privacy concerns, especially in sensitive indoor spaces such as bedrooms and bathrooms where many falls occur ( 32 ) . Such concerns significantly reduce user acceptance in real deployments ( 33 ) . To address these privacy issues, radio frequency (RF) sensing has recently emerged as a promising alternative ( 15 ; 64 ; 17 ) . RF-based systems enable device-free and non-visual monitoring, offering a more privacy-preserving solution. Despite their promise, existing RF-based fall detection approaches suffer from a fundamental limitation: their spatial resolution is insufficient to resolve fine-grained human body dynamics , resulting in elevated false alarm rates in real-world deployment. Most prior systems rely on coarse motion signatures, such as abrupt height changes ( 64 ) or velocity spikes ( 31 ) . These cues are easily confounded by fall-like daily activities ( e.g. , sitting down quickly or bending to pick up objects) and by fast-moving non-human entities such as pets or falling items ( 13 ) . In practice, even a seemingly modest false alarm rate of 2% ( e.g. , 1.8% reported in controlled evaluations ( 17 ) ) per one-minute monitoring window can accumulate into dozens of false alerts per day, overwhelming caregivers and eroding user trust—an outcome that critically undermines long-term deployment. To overcome these limitations, we revisit the fundamental biomechanical mechanism underlying a fall: the failure to maintain body balance . According to the World Health Organization, a fall occurs when an individual inadvertently comes to rest on a lower surface, typically following a transient yet decisive loss of balance ( 38 ) , as illustrated in Fig. 1 . Crucially, this loss of balance phase represents a distinct physical state in which the body’s posture exceeds its stability limits, rendering the subsequent descent inevitable. Therefore, reliable fall detection requires capturing body pose dynamics that govern balance states, rather than relying on indirect and coarse indicators such as height changes or velocity fluctuations commonly adopted in existing systems ( 64 ; 31 ) . Building on this insight, we present TaFall , a balance-informed fall detection system using a low-cost thermal array sensor. As shown in Fig. 1 , thermal arrays passively measure thermal radiation emitted by surrounding objects, including the human body, producing low-resolution temperature maps that inherently preserve privacy while retaining sufficient spatial structure to support pose dynamics estimation. Unlike camera-based systems, thermal arrays do not capture appearance-level details and operate robustly under darkness or low-light conditions typical of bedrooms and bathrooms ( 67 ) . Compared with RF modalities such as mmWave radar or WiFi sensing, thermal arrays provide finer spatial resolution 1 1 1 The thermal array sensor used in our prototype has a resolution of 62\times 80 within its field of view (FoV) 90^{\circ}\times 67^{\circ} , corresponding to angular resolutions of (1.13^{\circ},1.08^{\circ}) in azimuth and elevation, whereas a commonly used mmWave radar (TI IWR1843) provides (15^{\circ},60^{\circ}) resolution. and stronger human–background contrast due to the consistent human body temperature, thereby enabling more accurate reconstruction of pose dynamics that are essential for balance estimation. Leveraging these advantages, TaFall models a fall as a temporal progression of balance degradation . However, translating this idea into a practical fall detection system introduces significant challenges. ❶ Robust pose estimation. Thermal array sensors generate low-resolution ( e.g. , 62\times 80 for the sensor used in this paper), textureless temperature maps, as shown in Fig. 1 , making the recovery of fine-grained skeletal structures substantially more challenging than in RGB imagery. Moreover, rapid movements during a fall induce motion blur, which further obscures body contours and disrupts temporal consistency. These sensing characteristics jointly complicate stable extraction of balance-relevant pose dynamics from raw temperature maps. ❷ Body balance quantification. Body balance is commonly categorized into three stages, i.e. , stable balance (SB), loss of balance (LoB), and ground impact stage (GIS) ( 62 ; 65 ) , as illustrated in Fig. 1 (third row). Among these, the LoB phase is transient and subtle, making precise annotation difficult even with synchronized camera recordings. Furthermore, pose estimates derived from thermal data are sensitive to subject direction, distance, and environmental layout, introducing uncertainty into body balance state classification. These factors collectively challenge both ground-truth labeling and model-based inference of balance states.