Knee osteoarthritis is a chronic degenerative condition in which the cartilage that cushions the knee joint gradually wears away, producing pain, stiffness and progressively altered movement. It is one of the leading causes of disability in the United Kingdom, and its prevalence rises steeply with age. Because people with painful knees instinctively change how they walk — shortening strides, limping, shifting load to the healthier leg — gait carries a measurable signature of the disease. The challenge is capturing that signature reliably and affordably. The new survey argues that machine learning, and deep learning in particular, has reached a point where it can identify relationships in complex movement data that would be difficult or impossible for a human observer to detect, making automated gait analysis a realistic tool for detecting, diagnosing and monitoring musculoskeletal disorders with minimal human intervention.

At the heart of the survey lies a technical distinction that shapes the entire field: marker-based versus markerless human pose estimation. Pose estimation is the task of locating the human body in an image or video stream and identifying key anatomical points — hips, knees, ankles, shoulders — so that posture and movement can be quantified frame by frame. Marker-based systems, the traditional gold standard, require subjects to wear physical reflective or inertial markers placed at precise anatomical landmarks. Cameras or sensors track these markers with high accuracy, producing detailed three-dimensional kinematic data. The drawback is obvious: the procedure is time-consuming, requires specialist equipment and trained staff, and confines gait analysis to dedicated laboratories. For a condition like knee osteoarthritis, which affects millions and requires repeated monitoring over years, that model of care simply does not scale.

Markerless techniques flip the problem. Instead of tracking physical markers, deep neural networks learn to infer skeletal keypoints directly from ordinary video. Convolutional neural networks and, increasingly, transformer-based architectures are trained on vast annotated datasets of human movement, allowing them to estimate joint positions from a single camera view. The survey highlights how this opens the door to gait analysis in everyday environments — a smartphone camera, a webcam in a GP’s surgery — without any equipment attached to the patient. The trade-offs are equally clear. Markerless accuracy depends on lighting, clothing, camera angle and occlusion, and it typically yields two-dimensional or reconstructed three-dimensional estimates that are less precise than laboratory systems. For detecting the subtle gait deteriorations of early osteoarthritis, that precision gap matters, and the survey treats the comparison between the two paradigms as a central question rather than a settled one.

The most striking contribution of the review, however, is its attention to a third option that receives far less coverage in mainstream pose-estimation literature: thermal imaging. Thermal camera sensors capture the infrared radiation emitted naturally by the human body, producing images of heat distribution rather than reflected visible light. Because every person radiates heat, no external illumination is needed, and because thermal sensors record temperature patterns rather than identifiable facial features, they offer a non-invasive and privacy-preserving way to capture gait-related features. The survey examines how thermal data can complement or substitute for conventional video in pose estimation, potentially allowing movement analysis in complete darkness, in cluttered clinical settings, or in contexts where patients would reasonably object to being filmed with standard cameras.

The privacy argument deserves emphasis, because it is one of the reasons thermal imaging could prove decisive for real-world deployment. Standard video-based pose estimation raises genuine concerns: cameras in clinics, care homes or public spaces record faces, bodies and surroundings that patients may not consent to share, and the datasets used to train algorithms can expose sensitive information. Thermal imagery sidesteps much of this. A heat map of a walking person reveals the geometry of their stride and the temperature distribution of their limbs — which may itself carry clinical meaning, since inflammation and altered blood flow around an arthritic knee change local thermal patterns — while obscuring identity far more effectively than an RGB photograph. For a healthcare system that must balance innovation against data-protection obligations, a sensor that is simultaneously informative and privacy-preserving is an attractive proposition.

The survey situates these sensing technologies within the broader machinery of modern machine learning. Deep learning models excel at finding structure in high-dimensional data, and gait data is nothing if not high-dimensional: joint trajectories over time, spatiotemporal image sequences, pressure profiles and thermal signatures all must be fused and interpreted. Models trained on such data can classify gait patterns associated with musculoskeletal disorders, grade their severity and track progression over repeated assessments. The authors frame this as part of a wider revolution in which AI and machine learning have enabled systems to analyse complex patterns and make accurate predictions from large datasets across healthcare, detecting and monitoring conditions with minimal human intervention. Gait analysis, they argue, is a natural beneficiary: the signals are objective, quantifiable and rich, precisely the kind of input that deep networks handle well.

What emerges from the comparison is not a single winner but a trade-off space. Marker-based systems remain unmatched in accuracy and are likely to keep their place in research laboratories and biomechanical reference measurements. Markerless vision systems offer accessibility and scale, at the cost of some precision and robustness under uncontrolled conditions. Thermal imaging occupies a distinctive niche: it is non-invasive, works without lighting and protects privacy, but it is a younger technology in this application domain, with open questions about how well skeletal keypoints can be extracted from heat signatures alone and how large the training datasets are. The survey’s central message is that the choice of technique should follow the clinical context — screening a large population demands cheap, privacy-friendly, scalable sensing, while precise surgical planning may still justify the laboratory gold standard.

The clinical stakes are considerable. Early detection of knee osteoarthritis could shift management from reactive pain treatment toward proactive intervention — physiotherapy, weight management, lifestyle modification — before irreversible joint damage accumulates. Because gait changes often precede a patient’s decision to seek help, an automated, repeatable, low-cost measurement of how someone walks could serve as an early-warning system. The survey’s bridging of AI innovation with healthcare needs points toward exactly this: routine, unobtrusive gait monitoring integrated into ordinary care pathways, flagging deterioration for human clinicians to investigate. The authors are careful to present this as potential rather than accomplished fact; the field still needs validation studies, standardised datasets and regulatory-grade evidence before algorithms join the diagnostic toolkit.

Published open access in Artificial Intelligence Review, the survey arrives at a moment when the ingredients for this transformation — cheap cameras, powerful deep learning models and growing clinical interest in digital biomarkers — are all in place simultaneously. Its systematic comparison of marker-based, markerless and thermal approaches gives researchers a map of where the field stands and where the gaps lie, particularly around the underexplored promise of infrared sensing. If the authors are right, the future of musculoskeletal medicine may involve far fewer laboratories full of reflective markers and far more quiet, camera-free measurements of the most ordinary human act there is: putting one foot in front of the other. For the millions living with aching knees, a technology that watches how they walk — without ever really watching them — could change how early their disease is caught and how well it is managed.

Subject of Research: AI-based gait analysis and pose estimation techniques, including thermal imaging, for detecting knee osteoarthritis

Article Title: How does thermal imaging compare with other pose estimation techniques? A survey on gait analysis in healthcare applications with a focus on knee osteoarthritis

Article References: Abudayeh, M., Elizondo, D., Greenfield, S., & Caraffini, F. (2026). How does thermal imaging compare with other pose estimation techniques? A survey on gait analysis in healthcare applications with a focus on knee osteoarthritis. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11732-1

Image Credits: AI Generated

DOI: 10.1007/s10462-026-11732-1

Keywords: gait analysis, human pose estimation, thermal imaging, knee osteoarthritis, machine learning, deep learning, computer vision, healthcare AI, musculoskeletal disorders, markerless motion capture, privacy-preserving sensing, survey

Millimetre-wave radar, the same class of technology used in automotive collision avoidance and next-generation wireless communications, emits radio waves in the millimetre band and listens for reflections. When those reflections bounce off a moving person, the radar’s signal processing pipeline produces a point cloud: a scattered set of three-dimensional coordinates tracing the surfaces and limbs that returned the signal. Unlike the dense, well-structured point clouds produced by lidar or depth cameras, radar point clouds are notoriously sparse, often containing only a handful of points per frame, and they are contaminated by multipath reflections and hardware noise. That sparsity and irregularity is precisely what has made them so difficult for deep learning systems to interpret.

The Aston team’s insight begins with a widely used building block of point-cloud deep learning called EdgeConv, introduced originally in the Dynamic Graph CNN architecture. EdgeConv treats each point as a node in a graph, connects it to its k nearest neighbours, and computes features along the edges connecting them, capturing the local geometry of the scene. The problem, as the researchers point out, is that conventional EdgeConv is deterministic and indiscriminate: it always uses every one of its k nearest neighbours in every forward pass, faithfully incorporating noisy, redundant or uninformative points into the representation. On clean, dense benchmark datasets from 3D vision, that determinism is a virtue. On sparse, noisy radar data, it becomes a liability, because the same corrupted neighbourhoods poison the features pass after pass.

SEdgeNet’s central contribution is a redesigned operator the authors call Stochastic Edge Convolution, or SEdgeConv, which is built from two complementary mechanisms. The first, Stochastic Neighbour Sampling, randomly selects only a subset of each point’s k nearest neighbours during each forward pass, governed by a sampling ratio that determines what fraction of the neighbourhood is used. The idea echoes dropout-style regularisation techniques from the graph neural network literature, such as DropEdge: by forcing the network to produce useful representations from constantly varying, randomly drawn subsets of neighbours, the model is discouraged from leaning on any single noisy point or fixed local configuration. Crucially, the researchers note, this is not a permanent loss of information. Across the many forward passes of training, the full neighbourhood structure remains accessible; the stochasticity simply prevents overreliance on any particular slice of it.

The second mechanism, Edge-wise Factorised Convolution, tackles the computational side of the problem. Standard EdgeConv applies a dense transformation across all edge feature channels jointly, which is expensive. EFC instead adapts the depthwise separable convolution paradigm, originally developed to make convolutional networks efficient on regular image grids, to the irregular edge features produced by stochastically sampled neighbourhoods. Each channel of the edge feature tensor is first transformed independently, and the per-channel responses are then fused through a pointwise linear projection. The authors stress that the efficiency gain comes from coupling this factorised processing with the dynamically varying graph structure created by the stochastic sampling, rather than from the factorisation scheme alone. Neighbour features are then aggregated with symmetric max pooling, yielding a permutation-invariant representation that does not depend on the ordering of points.

Around these operators, SEdgeNet stacks multiple SEdgeConv layers and concatenates their outputs along the channel dimension, forming a multi-scale embedding that preserves both fine-grained local geometry and more abstract global structure. A dual pooling strategy, combining adaptive max pooling with adaptive average pooling, distils this embedding into a fixed-length global descriptor; max pooling emphasises salient structures while average pooling encodes overall shape statistics. The pooled vectors feed a three-layer multilayer perceptron classifier that outputs the predicted activity class. The model is trained with stochastic gradient descent, a cosine annealing learning rate schedule, cross-entropy loss and early stopping, with all reported results averaged over three independent training runs using different random seeds.

The team evaluated SEdgeNet on two public mmWave radar datasets. The first, MMActivity, contains 93 minutes of point cloud data covering five activities performed by two subjects, captured with a Texas Instruments IWR1443BOOST sensor and processed into 12,097 training and 3,538 test samples using two-second windows. The second, MiliPoint, is far more demanding: 49 activities performed by 11 subjects, recorded with a TI IWR1843 radar and split 80-10-10 for training, validation and testing. On MMActivity, SEdgeNet outperformed the best previous model by 0.72 percent in accuracy, 0.71 percent in precision, 2.7 percent in recall and 0.71 percent in F1 score. On MiliPoint, the gains were larger still: 5.42 percent in accuracy, 4.91 percent in precision, 4.81 percent in recall and 5.26 percent in F1 score, establishing new state-of-the-art results on both benchmarks.

What makes those numbers remarkable is that they were achieved while using fewer computational resources than every baseline compared. SEdgeNet has the fewest parameters and the lowest multiply-accumulate count among the tested methods. Against DGCNN, the most efficient baseline, it uses 0.12 million fewer parameters and 0.02 giga-MACs less computation while achieving dramatically higher accuracy. Against AGConv, the most accurate baseline, it improves accuracy by 5.42 percent while cutting parameters by 0.52 million and computation by 8.43 giga-MACs. In controlled experiments swapping SEdgeConv for standard EdgeConv within an otherwise identical architecture, the stochastic version improved accuracy by 9.4 percent on MiliPoint and 0.44 percent on MMActivity, and required 3.08 giga-MACs less computation at a neighbourhood size of five. Ablation studies confirmed that both SNS and EFC contribute independently, and that combining them yields the best accuracy on both datasets alongside the lowest computational cost of 0.34 giga-MACs.

The hyperparameter analysis revealed instructive differences between the two datasets. MMActivity, with only five relatively separable classes, peaked with two layers and 35 nearest neighbours, reaching 98.12 percent accuracy, and its performance was dominated by network depth rather than neighbourhood size. MiliPoint, with 49 harder-to-distinguish classes, required a larger neighbourhood of 50 neighbours and five layers to reach 85.70 percent, suggesting that harder classification tasks demand both wider receptive fields and deeper networks to extract sufficiently discriminative geometric structure. The optimal sampling ratio also differed, at 0.6 for MMActivity and 0.3 for MiliPoint, though the authors found that a range of 0.3 to 0.5 delivers near-optimal accuracy on both datasets and recommend it as a practical default. Because the stochastic sampling remains active at inference time, the team also examined a Monte Carlo-style deployment strategy: averaging predictions over multiple forward passes. On MMActivity, accuracy rose from 97.72 percent with a single pass to a peak of 98.34 percent at six passes, at a still-modest cost of 2.04 giga-MACs.

Perhaps most compelling is the demonstration that the system runs in real time on embedded hardware. The researchers deployed their trained pipeline on a Jetson Nano mounted on a Robotino robot, with an AWR1843BOOST radar feeding data over USB. In a proof-of-concept trial, a single subject performed sitting, standing and walking freely in an indoor room, and the system achieved 99 percent classification accuracy, recognising actions under both lit and dark conditions. Average preprocessing latency was 12.18 milliseconds and model inference averaged 15.14 milliseconds, for a combined end-to-end latency of 27.32 milliseconds, comfortably within real-time requirements. The authors are careful to frame this as a feasibility demonstration rather than a comprehensive deployment study, noting that broader evaluation across multiple subjects, more activities and varying distances remains future work. Even so, the combination of privacy preservation, darkness-robust sensing, state-of-the-art accuracy and embedded real-time performance positions SEdgeNet as a promising foundation for assistive robotics, ambient intelligence and independent-living support, where the goal is to watch over people’s wellbeing without ever truly watching them.

Subject of Research: Graph neural network for human activity recognition from sparse millimetre-wave radar point clouds

Article Title: SEdgeNet: stochastic edge network for human activity recognition using sparse point cloud

Article References: Zakka, V. G., Manso, L. J., & Dai, Z. (2026). SEdgeNet: stochastic edge network for human activity recognition using sparse point cloud. Journal of Ambient Intelligence and Humanized Computing. https://doi.org/10.1007/s12652-026-05133-z

Image Credits: AI Generated

DOI: 10.1007/s12652-026-05133-z

Keywords: SEdgeNet, millimetre-wave radar, human activity recognition, point cloud, graph neural network, stochastic neighbour sampling, edge convolution, privacy-preserving sensing, embedded AI, ambient intelligence, depthwise separable convolution, assistive robotics