Oxygen uptake ($\dot{\mathrm{V}}\mathrm{O}$${}_2$) is an important metric in any exercise test including walking and running. It can be measured using portable spirometers or metabolic analyzers. Those devices are, however, not suitable for constant use by consumers due to their costs, difficulty of operation and their intervening in the physical integrity of their users. Therefore, it is important to develop approaches for the indirect estimation of $\dot{\mathrm{V}}\mathrm{O}$${}_2$-based measurements of motion parameters, heart rate data and application-specific measurements from consumer-grade sensors. Typically, these approaches are based on linear regression models or neural networks. This study investigates how motion data contribute to $\dot{\mathrm{V}}\mathrm{O}$${}_2$ estimation accuracy during unconstrained running and walking. The results suggest that a long short term memory (LSTM) neural network can predict oxygen consumption with an accuracy of 2.49 mL/min/kg (95% limits of agreement) based only on speed, speed change, cadence and vertical oscillation measurements from an inertial navigation system combined with a Global Positioning System (INS/GPS) device developed by our group, worn on the torso. Combining motion data and heart rate data can significantly improve the $\dot{\mathrm{V}}\mathrm{O}$${}_2$ estimation resulting in approximately 1.7–1.9 times smaller prediction errors than using only motion or heart rate data.

Radar-based human activity recognition is a promising privacy-preserving alternative to vision and wearable sensors, especially in sensitive environments such as long-term care facilities. However, reliable detection of quasi-static human presence, where subtle micro-motions such as small postural sway, breathing, small limb readjustments, etc, concentrate their energy around near-zero Doppler using low-resolution radar, remains challenging. In this work, we propose Range-Angle based Semi-Static Occupancy (RASSO), an invertible Doppler-domain non-linear remapping layer that resamples the slow-time FFT (Doppler) axis to densify the sampling grid around near-zero Doppler before adaptive Capon beamforming. This produces Range-Azimuth (RA) feature maps with higher Signal-to-Noise ratios (SNR), resulting in sharpened target peaks and reduced background variance. On a real dataset collected in a nursing home using short-range SIMO FMCW radar, RASSO consistently yields higher-contrast RA maps that are easier to threshold and more informative for downstream models. Under cell-averaging constant false alarm rate (CA-CFAR), RASSO-enhanced RA (RASSO-RA) achieves the highest detection-curve area under the curve (AUC = 0.981) and the best recall at false alarm rates (FAR) of 1% and 5% (0.920 and 0.947), compared to conventional Capon beam formation and a recent state-of-the-art baseline. Training on RASSO-RA further improves data-driven methods: a frame-based SimpleCNN reaches 95–99% accuracy across subjects, while a sequence-based CNN–LSTM achieves 99.4–99.6% accuracy, with paired session-level bootstrap analysis confirming statistically significant macro-F1 improvements of 2.6–3.6 percentage points (95% confidence intervals strictly above zero) over the non-warped baseline.

Oxygen consumption ( V ˙ O 2 ) is an important measure for exercise test, such as walking and running, that can be measured outdoors using portable spirometers or metabolic analyzers. However, these devices are not feasible for regular use by consumers as they intervene with the user’s physical integrity, and are expensive and difficult to operate. To circumvent these drawbacks, indirect estimation of V ˙ O 2 using neural networks combined with motion features and heart rate measurements collected with consumer-grade sensors has been shown to yield reasonably accurate V ˙ O 2 for intra-subject estimation. However, estimating V ˙ O 2 with neural networks trained with data from other individuals than the user, known as inter-subject estimation, remains an open problem. In this paper, five types of neural network architectures were tested in various configurations for inter-subject V ˙ O 2 estimation. To analyse predictive performance, data from 16 participants walking and running at speeds between 1.0 m/s and 3.3 m/s were used. The most promising approach was Xception network, which yielded average estimation errors as low as 2.43 ml×min⁻¹×kg⁻¹, suggesting that it could be used by athletes and running enthusiasts for monitoring their oxygen consumption over time to detect changes in their movement economy.

Oxygen consumption is an important parameter for exercise test, such as walking and running, that can be measured using portable spirometers or metabolic analyzers. However, these devices are not feasible for regular use by consumers as they intervene with the user’s physical integrity, and are expensive and difficult to operate. To circumvent these drawbacks, indirect estimation of using neural networks combined with motion parameters and heart rate measurements collected with consumer-grade sensors has been shown to yield reasonably accurate for intra-subject estimation. However, estimating with neural networks trained with data from other individuals than the user, known as inter-subject estimation, remains an open problem. In this paper, five types of neural network were tested in various configurations for inter-subject estimation. To analyse predictive performance, data from 16 participants walking and running at speeds between 1.0 m/s and 3.3 m/s were used. The most promising approach was XceptionNet, which in most configurations even yielded a lower average estimation error than the LSTM neural network from an earlier study for intra-subject estimation. This suggests that XceptionNet could be embedded in portable devices for real-time estimation of oxygen consumption during walking and running.