Bicycle-rider control identification

Abstract

Statistics show that cycling accidents have the biggest (and increasing) contribution to the overall number of hospital visits related to traffic accidents in the Netherlands. In the majority of these cycling accidents, no other road users are involved. Because the majority of the cycling accidents are so-called single vehicle accidents, understanding the control behaviour of the cyclist can spark new insight in to effective preventive measures. This thesis aims to achieve that by reviewing the current state-of-the-art in bicycle-rider control research and subsequently proposing a new rider control model that is developed using system identification techniques. The results of a literature review are used to propose a novel bicycle-rider control model structure that includes realistic human neural and cognitive characteristics to predict control behaviour for stabilizing a bicycle. Sensory integration is modelled with a Kalman filter, and human control and prediction capabilities are mimicked with a Linear-Quadratic-Regulator (LQR) and Tapped-Delay-Line predictor, respectively. Human neuromuscular dynamics are included in the model structure in the form of a second order filter and the bicycle dynamics are modelled with the linear Whipple bicycle model. Two different experimental data sets during which cyclists are laterally (roll) perturbed on instrumented bicycles are used: one collected at the UC Davis on both a horse treadmill and in a sports pavilion and the other collected at the TU Delft on a public cycling path (without other road users). The human steer response to the perturbation is separated from other effects caused by unknown disturbances and noise with a non-parametric Finite-Impulse-Response model. A structured system identification approach is used to determine the final free parameter set that approximates the non-parametric model in the best way possible. The data sets are split in a train set and a test set. The model structure is fitted to the train set. The predictive performance is verified by applying the resulting model to the test set. For the UC Davis data, the resulting model has one rider dependent parameter that describes human bicycle-balancing control behaviour for the entire evaluated forward velocity range (2-8 m/s). This parameter is the weight placed on the roll angle in the LQR control algorithm. The single-run training performance in terms of Variance-Accounted-For (VAF) with the non-parametric model reaches up to 97 %. Test performance incurred a VAF drop, on average, of around 10 %. The steer responses of the TU Delft experiment can be predicted with a model that has two rider dependent parameters: the weights placed on the roll angle and roll rate in LQR control algorithm. Training performance reaches up till 83 % and test performance incurred a VAF drop, on average, of around 7%. Promising future directions are the inclusion of a passive rider model to further increase the fidelity of the model. The investigation of the stability margins of the proposed bicycle-rider controller and their sensitivity to bicycle design and/or human sensory and cognitive decline can lead to tailored measures to reduce the injury rate among cyclists.