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.
Figures & tables
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 .
Figure 2. Thermal array sensing characteristics and biomechanical context. (a) Operating principle of a thermal array sensor, which captures thermal radiation emitted by the human body to generate a low-resolution temperature map. The standing subject is annotated with the Center of Mass (CoM) and Base of Support (BoS). (b) Comparison of static and dynamic postures under varying motion velocities. Thermal arrays produce stable temperature maps under static conditions, whereas rapid motion induces pronounced thermal motion blur (highlighted by white boxes), a phenomenon largely absent in RGB imagery. The progressive expansion of the highlighted regions indicates that motion blur magnitude increases with movement velocity.
Figure 3. Overview of the TaFall framework. TaFall consists of two complementary phases. (a) Balance-Informed Fall Detection : an Appearance–Motion Fusion module estimates a robust 2.5D human pose sequence from thermal array temperature maps by jointly leveraging spatial appearance cues and motion-blur–induced dynamics. The resulting pose sequence is then processed by a Balance-Aware Pose Network , which learns a Physically grounded balance representation to jointly infer continuous balance states and detect fall events. (b) Pose-Bridged OOV Enhancement : the balance-aware pose network is pretrained using large-scale public pose data projected into diverse 2.5D views, with cross-view contrastive learning enforcing motion-sensitive and viewpoint-invariant representations, thereby enhancing robustness and reducing false alarms in real-world deployments.
Figure 4. The structure of the Spatial Branch. It first predicts the Center Heatmap and the Human Scale Map from the thermal sequence. The two outputs are then combined to generate a Masked Temp Map, which is further processed by convolution layers to extract the Appearance Feature .
Figure 5. The structure of Motion Blur Branch. It first generates the Motion History Image and then uses an attention mechanism to extract motion feature from temperature map and motion history image.
Figure 6. The illustration of the Physically Grounded Balance Representation and the Balance-Aware Pose Network. (a) Physically Grounded Balance Representation : Balance is quantified using the Signed Margin of Balance (SMoB), defined as the signed distance between the projected center of mass (CoM) and the boundary of the base of support (BoS). (b) Balance-Aware Pose Network : The predicted pose sequence is encoded using ST-GCN-based pose encoder to capture postural evolution. The encoded pose representation is then fed into two lightweight prediction branches: a temporal transformer head for fall-event classification, and an MLP head for SMoB prediction.
Figure 7. Illustration of the Pose-Bridged OOV Enhancement strategy. A motion-capture dataset provides 3D skeletal trajectories and is converted into the 2.5D coordinate space via 3D-to-2.5D Random Projection by sampling virtual camera viewpoints, producing multiple 2.5D projected views of the same motion sequence. The projected pose sequences are used to pretrain the Balance-Aware Pose Network with a multi-objective formulation that combines fall classification, balance regression, and an additional View-Consistent Contrastive Learning objective to encourage cross-view feature consistency.
Figure 8. Experimental setup. Fig. (a) illustrates the three representative indoor environments used for fall data collection. Fig. (b) shows the Meridian MI0802M6S thermal array sensor deployed in all sessions. Fig. (c) presents the four real-world scenarios for long-term evaluation.
Figure 9Figure 10Figure 11Figure 12
Figure 19. Long-term false alarm case study in real environment. Fig. (a) shows five types of false alarms(FA), where the red boxes indicate the bounding boxes predicted by the detection model of TaFall , Fig. (b) shows the number of different false alarm types together with the total number of samples.
Figure 20. Bathroom case study. (a) Examples of collected daily activities and falls in bathroom environment. (b) False alarm example in a bathroom environment caused by thermal accumulation on the floor when the subject is occluded by the wall. (c) Temperature map during showering in bathroom, where thermal reflections from falling water do not interfere with pose estimation or fall detection.