Truck Lateral Velocity Estimation Using Deep Machine Learning

Background of thesis project

One of Volvo GTT’s prime goals going into the next generation of vehicles is automation. To autonomously control these vehicles in a safe manner puts tough demands on accurately knowing the current dynamic state of the complete vehicle combination, e.g., velocities, accelerations and rotational velocities of both truck and semi-trailer. In addition to the automation aspect, it's very valuable to know the vehicle's state from a pure safety perspective also under manual driving. If we know the dynamics of the vehicle, we can, e.g., prevent accidents such as roll-over, jack-knifing and trailer swingout.

Although the vehicle has sensors such that many of these states are directly observed, perhaps the most important state to understand the vehicle’s dynamic performance, the lateral velocity (see Fig 1), is typically not. Furthermore, its relation to measured quantities is highly dependent on unknown parameters, such as vehicle mass, road friction and tire stiffness. To make the problem even more challenging, for safe automation it is crucial to know how precise our current estimates are. That is, that our estimator can with high precision describe its uncertainty.

Estimation of the lateral velocity has previously been explored for cars, but the problem is much less studied for trucks. Compared to cars, trucks experience a larger lateral force and having multiple axles complicate the model even further.

Suitable background

Strong background in machine learning and Python with a good mathematical background.

Description of thesis work

This thesis aims to develop and investigating advanced machine-learning methods to solve problems within autonomous driving. More precisely, the focus is on developing machine-learning-based methods to estimate the lateral velocities (v1y and v2y) of the tractor and semitrailer, respectively, using onboard sensors. The emphasis of the estimator should be put on, in addition to providing accurate estimates, also providing a precise description of the estimation uncertainty. Your proposed method will be trained and evaluated on both annotated simulated and real data that is readily available.

One possible way to approach the problem is to construct an ensemble of neural networks with additional randomized prior functions (RPF) [1]. For these networks, each individual network is trained to produce an independent prediction and, an uncertainty estimate can be found by considering the spread of these individual predictions. However, there are many other ways of approaching this problem and this is what will be explored in the thesis.

The work will include literature survey, choosing, implementing and adapting promising ML approaches to your problem and evaluating the result on relevant use cases using both real and simulated data.

This project is one of two, where the other one will try to solve the problem using a model-based filter. The work will be carried out at Volvo Group Trucks Technology, Sweden.

Thesis Level: Master

Language: English

Starting date: January 2023

Number of students: The thesis is recommended for one or two students

Tutors:

Axel Ceder – Chalmers / Volvo GTT

mail: [email protected]

phone: +46 73 076 59 50

Mats Jonasson – Chalmers / Volvo GTT

mail:[email protected]

phone: +46 03 491 373