Vehicle Collision Avoidance in NLoS Scenarios Using Machine Learning-Assisted Positioning and MQTT

Abstract

Intersections have long been recognized as high-risk areas for traffic accidents, with vehicle collisions in Non-Line-of-Sight (NLoS) scenarios resulting in significant economic
and human losses. This thesis investigates a vehicle collision avoidance system that integrates Machine Learning (ML), wireless communication technologies, and sensor fusion
with Global Navigation Satellite System (GNSS) positioning, aiming to bridge existing
gaps in current research by incorporating the latest technologies and validating their effectiveness in real-world scenarios.
In this study, various technologies are employed, including a combination of RealTime Kinematic (RTK) GNSS and dead reckoning methods to achieve high-precision
vehicle localization. A Fully Connected Network (FCN) model is utilized to predict future
trajectories with an error margin of two to three meters over a three-second prediction.
Additionally, Message Queuing Telemetry Transport (MQTT) technology is employed to
facilitate wireless communication between vehicles.
Experimental results demonstrate that our system effectively provides collision warnings and prevents vehicle collisions in both Line-of-Sight (LoS) and NLoS scenarios.
The findings of this research illustrate the substantial potential of combining wireless
communication technologies and ML in advancing vehicle collision avoidance, establishing it as an effective tool for enhancing traffic safety. Future work will focus on optimizing
trajectory prediction accuracy and exploring system performance across a wider range of
real-world scenarios.

Popular Abstract

Have you ever worried about a potential collision when driving through an intersection,
especially when you can not see the other vehicles coming? Thanks to advancements
in technology, we can now use Machine Learning (ML) to predict vehicle movements
and avoid crashes before they happen. This thesis explores how we can improve vehicle
positioning accuracy and predict trajectories using a combination of Global Navigation
Satellite System (GNSS), Real-Time Kinematic (RTK) corrections, and ML.
GNSS has been a standard for vehicle positioning, but its accuracy is usually limited
to about three to five meters. This level of precision is not enough for collision avoidance,
especially in crowded or complex environments. By combining GNSS with RTK, we can
improve the accuracy to within a few centimeters. RTK works by correcting the GNSS
signals with data from nearby reference stations. This is important because GNSS signals
are affected by environmental factors like tall buildings or weather, which can introduce
errors. RTK uses the known position of these reference stations to reduce this error, giving
us a much more precise location for vehicles.
Once we have this accurate position, the next step is to predict where the vehicle
will be in the next few seconds. There are a lot of sensors that have been integrated
into the vehicle such as camera, Light Detection and Ranging (LiDAR), gyroscopes, and
accelerometers. A lot of information could be gathered from those sensors. This fusion
of data, known as sensor fusion, helps estimate things like speed, direction, and distance.
Dead reckoning, another technique used in navigation, helps estimate the vehicle’s current
position based on previous data. However, dead reckoning tends to accumulate errors over
time, which can lead to inaccurate predictions. A traditional Kalman filter could reduce
the error but it does not solve the accumulated error problem. This is where ML comes
into play. The advantage of ML is that it can learn from these errors. By recognizing
patterns in how errors are generated, the model can compensate for them, allowing us
to predict vehicle positions with greater accuracy. This predictive ability is crucial for
preventing collisions, as it gives vehicles time to adjust their path or warn drivers before
an accident occurs.

• author: He, Yumei and Guan, Yuqin
• supervisor: Xuesong Cai; Meifang Zhu
• organization: Department of Electrical and Information Technology
• course: EITM02 20241
• year: 2024
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• report number: LU/LTH-EIT 2024-1023
• language: English
• id: 9176401
• date added to LUP: 2024-10-16 10:43:25
• date last changed: 2024-10-16 10:43:25