Inertial navigation has made numerous applications for many years. One of the leading advantages is the self-contained ability to track the position and orientation of a vehicle. Nowadays, micro-electro-mechanical systems (MEMS) inertial measurement unit (IMU) enable the users to have more opportunities developing more dead-reckoning (DR) techniques on the small chip and build a great amount of potential in wheeled platform market, such as shared bike, shared e-scooter, automobile, etc. However, the conventional DR solution is not adaptive for different wheeled platforms, especially in two-wheeler applications.

The abnormal or nonlinear motion usually makes the conventional DR drift or violate the vehicle motion constraints, such as hanging the bike or rotating the handlebar without moving, and it often occurs in real life scenario. This paper proposes the new DR algorithm to widely support not the traditional automobile but also two-wheeler platform (bike and e-scooter). The contributions of this paper are given as follows.

(1) To adaptively enhance the DR solution, the machine learning (ML)-based vehicle motion detection is implemented as the pseudo-measurement to constrain the error of DR. In this paper we minimize the performance degradation specifically due to the abnormal motion. The proposed ML-based DR algorithm utilizes the decision tree (DT) to identify the abnormal motions of the vehicle and develops corresponding motion constrains to overcome the GNSS outage, multipath situations. The proposed DT model uses not only the pure IMU data as the features, but also the features converting into time domain and spatial domain to increase the accuracy of the recognition. The proposed DT model can recognize the nonlinear motions of the vehicle, preventing the DR algorithm from using vehicle motion model, and applies the corresponding motion constraints properly.

(2) In other words, the proposed DR algorithm adapts the correct and suitable vehicle motion constraints to enhance the performance. We propose the linear virtual vehicle motion speed model when the DT model detects the user is in the linear moving situation. It is one of the important features as the odometer-like measurement to control the position and velocity drift.

(3) Finally, the proposed ML-based DR algorithm is implemented on the Airoha Technology (a subsidiary of MediaTek Group) AG3335 series GNSS chipset. The sensors employed here are low-cost, consumer-grade 6-axis IMUs (it combines a 3-axis gyroscope, and a 3-axis accelerometer) and largely available in commercial vehicles. This paper evaluates the proposed method using the real-world data collecting around the urban area. The results presenting in this paper include the bike and e-scooter data in different scenarios and with different motions. During the experiments, we simulated the user behaviors which are nonlinear motions during the movement.

Moreover, most of the designed trajectories are close to the buildings to evaluate how well the proposed DR solution. The preliminary results illustrate that proposed algorithm performs better than traditional DR solution or GNSS solution. It is noted that the proposed DR solution can address the nonlinear motions and improves the positioning accuracy by mitigating the multipath. By integrating machine learning-based motion detection to identify and alleviate the abnormal activities, the horizontal positioning accuracy over this mechanism can be enhanced to 50% in urban area, and 30% in vertical positioning accuracy typically.

Learn more on https://www.ion.org/gnss/virtual-abstract-view.cfm?paperID=11240