In this article, a statistical model of human motion as observed by a network of radar sensors is presented where knowledge on the position and heading of the target provides information on the observation conditions of each sensor node. Sequences of motions are estimated from measurements of instantaneous Doppler frequency, which captures informative micromotions exhibited by the human target. A closed-form Bayesian estimation algorithm is presented that jointly estimates the state of the target and its exhibited motion class which are described by a hidden Markov model. To correct errors in the estimated motion class distribution introduced by faulty modeling assumptions, calibration of the probability distribution and measurement likelihood is performed by isotonic regression. It is shown, by modeling sensor observation conditions and by isotonic calibration of the measurement likelihood that a cognitive resource management system is able to increase classification accuracy by 5%-10% while utilizing sensor resources in accordance with defined mission objectives.

@en

Cognitive radar frameworks rely on the ability to quantify and reason on future uncertainty, which allows for the selection of an optimal decision policy. These methods require that the uncertainty estimates provided by the underlying statistical model are well-calibrated, i.e. consistent with true uncertainty. In this work, the utilization of probability calibration techniques for target classification is explored. It is shown from simulations and experimental data that the proposed techniques can be used to correct errors in uncertainty estimates caused by incorrect modeling assumptions, such as the independence of sensors and the independence of classification covariates. This correction improves classification performance and the reliability of cognitive systems so that resources are utilized in accordance with user-defined cost functions.@en

In this paper, the classification of human activity from micro-Doppler spectrograms measured by a radar network is considered. To cope with differences between the training and test datasets due to changes in the set of participants, signal-to-noise ratio and polarimetry, domain adaptation is proposed. To realize this, linear mapping between the two domains is assumed and estimated by one of two methods, expectation-maximization or empirical estimates of statistical moments. The performance of the methods is evaluated on experimental data measured by a multi-static radar network. The proposed methods increase the classification accuracy by 5–15 percentiles on the recorded dataset.@en

Perception systems for autonomous vehicles are reliant on a comprehensive sensor suite to identify objects in the environment. While object recognition systems in the LiDAR and camera modalities are reaching maturity, recognition models on sparse radar point measurements have remained an open research challenge. An object recognition model is here presented which imposes a graph structure on the radar point-cloud by connecting spatially proximal points and extracts local patterns by performing convolutional operations across the graph’s edges. The model’s performance is evaluated by the nuScenes benchmark and is the first radar object recognition model evaluated on the dataset. The results show that end-to-end deep learning solutions for object recognition in the radar domain are viable but currently not competitive with solutions based on LiDAR data.@en