Trajectory Prediction for Powered Two Wheelers with Deep Learning

Type of the Paper: Extended Abstract

Trajectory Prediction for Powered Two Wheelers with Deep Learning

Karl Ludwig Stolle^1,*, Anja Wahl¹, and Stephan Schmidt²

¹Robert Bosch GmbH; karlludwig.stolle@bosch.com, https://orcid.org/0000-0002-6535-1894; anja.wahl@bosch.com

²University of Applied Science Merseburg

*corresponding author.

Name of Editor: Jason Moore

Submitted: 28/02/2023

Accepted: 16/04/2023

Published: 26/04/2023

Citation: Stolle, K., Wahl, A. & Schmidt, S. (2023). Trajectory Prediction for Powered Two Wheelers with Deep Learning. The Evolving Scholar - BMD 2023, 5th Edition.

This work is licensed under a Creative Commons Attribution License (CC-BY).

Abstract:

The development of active safety systems for powered two wheelers (PTW) plays an important role to reduce the number of accidents involving PTW. Three quarter of PTW accidents in Europe are non-single accidents (Brown et al., 2021), in which riders are prone to heavy injuries or fatalities due to inherently low passive safety of PTW. The Connected Motorcycle Consortium (CMC) investigates the use of vehicle-to-X (V2X) communication to avoid collisions by warning the rider and drivers of surrounding vehicles in case of an imminent risk of collision (Connected Motorcycle Consortium, n.d.). Beside collisions, single vehicle accidents make up of 25 % of PTW accidents with fatalities or serious injured riders, 64 % of them occur during cornering (Brown et al., 2021). The majority of single accidents are primarily caused by rider error and could therefore be avoided (Biral et al., 2014). Future Advanced Rider Assistance Systems (ARAS) addressing either the V2X or the single vehicle accident domain will require information about the riding intention. The calculation of a trajectory in the upcoming seconds of the ride is one way to describe this intention. The presented work pursues the prediction of the PTW lateral dynamic state by means of a roll angle trajectory over the upcoming 4 s of riding. Inputs for the prediction model are PTW internal signals only, that are measurements of vehicle dynamics, rider inputs, and rider behavior.

Making a prediction based on any physical model of a PTW is limited to the time constant between rider inputs and the dynamic states. A previous study of the authors revealed that the time delay between the first steer torque input and the roll angle state ranges between 0.45 s and 2 s for the given test motorcycle, dependent on velocity (Stolle et al., 2022). Consequently, empirical methods need to be applied to achieve further predictions. Scherer and Basten (2022) present a parameterizable mathematical model for the prediction rider-individual and curve-individual roll angle trajectories. Upon reviewing the on-road riding data to be used in this work, it became apparent that the effort required to develop a heuristic model for the prediction is unforeseeable high, coupled with uncertain chances of success. Therefore, it was decided to use the exploratory capability of a deep learning (DL) model to investigate the possibilities in predicting the PTW roll angle state based on the given riding data. An ablation study is realized to understand the feature importance of the non-common measurement signals of rider steering inputs and rider behavior.

On-road data of more than 70 h ridden by 21 riders on an equipped test motorcycle is available for the development of the DL model. It contains a broad variety of routes from all over Southwest Germany with a focus on rural roads. Only riding with a velocity > 30 km/h is regarded for the prediction task to ensure the motorcycle is in the stable regime: this leaves ~ 65 h of data remaining. Beside filtering of signals, the riding data needs further preparation. Removing the bias from an outweighing share of straight riding condenses the overall data to ~ 28 h used for training, validation, and test.

Figure 1. Drawing of the DL model’s time-series prediction task: multiple features F are used to predict P at current time ti

Concrete task of the DL model is the prediction of future roll angle values P from the current point in time t_i up to a certain maximum preview time t_i+P at multiple discrete points that are evenly spaced with T_P – this prediction vector will be referred to as prediction horizon. Input to the model is a set of n features F, all time-series data, from t_i-H in the past up to the current point in time t_i at discrete points sampled with T_H. The prediction task described is visualized schematically in Figure 1. The DL model consists of one long short-term memory (LSTM) layer that is followed by a multilayer perceptron (MLP), which is a common neural network architecture for time-series prediction (Altche & De La Fortelle, 2017). Hyperparameter optimization is done on the network’s size, training parameters, and sampling T_H and history length t_i-H of the input features.

The prediction performance of the optimized DL model on test data is presented as “best model” in Figure 2, where the left graph shows the root mean square error (RMSE) of the roll angle prediction for each point along the 4 s prediction horizon separately. This DL model uses a set of 16 features including non-common sensors for steering torque, rate & angle (“steering system”) as well as for rider upper body lean & offset and rider head yaw angle (“rider behavior”). A basic prediction approach assuming constant roll angle over the prediction horizon is evaluated on the same test data and presented in the left graph of Figure 2 to contextualize the DL model’s performance; its overall RMSE is 54 % higher, as shown in the table on the right of Figure 2.

Furthermore, a detailed ablation study on the importance of the non-common input features is carried out. Its condensed results are presented in the table on the right of Figure 2 as the change in overall RMSE loss compared to the best model for two different input configurations. Removing the rider behavior features in the “standard + steering system” configuration increases the error by 5 %. Additionally removing the steering system features leaves only common measurement signals in the “standard” configuration behind and leads the RMSE to increase by 7 % over the full “standard + steering system + rider behavior” model.

In summary, the LSTM based DL model predicting a roll angle trajectory is a promising approach for the detection of the riding intention regarding lateral dynamics. The method demonstrates that there is information in the history of the time-series measurement signals of a PTW that is valuable for the prediction of future roll angles. The results of an ablation study reveal the importance of different input features and thus highlight the positive effect of non-common measurement signals, as the removal of both steering system and rider behavior measurements decreases the overall prediction performance. Future work will investigate how maneuvers and riding styles affect the prediction.

References

Altche, F., & De La Fortelle, A. (2017). An LSTM network for highway trajectory prediction. In IEEE 20th International Conference on Intelligent Transportation Systems (ITSC) (pp. 353–359). IEEE.

Biral, F., Bosetti, P., & Lot, R. (2014). Experimental evaluation of a system for assisting motorcyclists to safely ride road bends. European Transport Research Review, 6(4), 411–423. https://doi.org/10.1007/s12544-014-0140-6

Brown, L., Morris, A., Thomas, P., Ekambaram, K., Margaritis, D., Davidse, R., Usami, D. S., Robibaro, M., Persia, L., Buttler, I., Ziakopoulos, A., Theofilatos, A., Yannis, G., Martin, A., & Wadji, F. (2021). Investigation of accidents involving powered two wheelers and bicycles - A European in-depth study. Journal of Safety Research, 76, 135–145. https://doi.org/10.1016/j.jsr.2020.12.015

Connected Motorcycle Consortium. (n.d.). Applications to improve rider safety. Retrieved February 3, 2023, from https://www.cmc-info.net/applications.html

Scherer, F., & Basten, T. (2022). Entwicklung eines Motorradfahrendenmodells zur Trajektorienprädiktion. In Insitut für Zweiradsicherheit (ifz) e.V. (Ed.), ifz-Research Publication Series: Nr. 20, Sicherheit - Umwelt - Zukunft XIV: Tagungsband der 14. Internationalen Motorradkonferenz 2022 (pp. 265–299). Institut für Zweiradsicherheit.

Stolle, K. L., Wahl, A., & Schmidt, S. (2022). Importance of motorcycle rider upper body movement for rider intention detection and motorcycle state prediction. In Insitut für Zweiradsicherheit (ifz) e.V. (Ed.), ifz-Research Publication Series: Nr. 20, Sicherheit - Umwelt - Zukunft XIV: Tagungsband der 14. Internationalen Motorradkonferenz 2022 (pp. 32–41). Institut für Zweiradsicherheit.