Master's student awarded for research on traffic flow patterns

Akshay Gurudath receives the Christer Gilén Scholarship for his degree project in statistics and machine learning. In his thesis, Akshay presents and compares models he has built to predict positions of pedestrians, cyclists, and autonomous vehicles.

Akshay Gurudath at the white board Akshay Gurudath

“I think it was a mix of both surprise and gratitude”, says Akshay Gurudath about his initial reaction to the prize for his thesis entitled “AI-based prediction of road users’ intents and reactions”.
 
Akshay arrived at Linköping University and the Department of Computer and Information Science (IDA) from southern India in September 2020.
 
“I had read a little bit of the research they had done, and it was very interesting. At the same time, I really liked the master’s programme in statistics and machine learning because it had a mix of both statistics and applications of statistics, which not many programmes have, and that balance between theory and implementation was what got me hooked on to Linköping.”

How did you choose the topic for your thesis?

The degree project was done in collaboration with the startup company Viscando in Gothenburg.

“I was initially excited about what the company does. They were measuring traffic information and they were also trying to gauge movement of cyclists or pedestrians in crowded intersections, and they would present their analysis of traffic flow to different stakeholders such as municipalities and businesses. They did this with sensors based on 3D and artificial intelligence (AI) technology that they built and developed.”

Akshay sent a long email to one of the heads of the company and was invited to Gothenburg for an interview. At the interview, they discussed the field of measuring intent of pedestrians and subsequently decided to take on this project of trying to predict intent, at least to a degree, by using a person’s past trajectory.

The objective was also to observe events during slow traffic conditions. Akshay explains that autonomous driving is becoming more and more prevalent in the world. However, self-driving cars are usually studied in faster driving conditions where there is more at stake.

If we were to have autonomous delivery robots in the future, that deliver food for example, we wanted to see how these autonomous robots can move around in a slow traffic environment.


To minimise the interruption to traffic by such robots, it will be important to understand how pedestrians and cyclists move.

Another motivation and more immediate reason for the project was a need for the company to better understand movement patterns in warehouses to minimise accidents involving vehicles and workers.

What is machine learning?

“I think there are many different definitions of machine learning. I like to think of it as follows. There are usually patterns associated with any problem/data and machine learning is about understanding these patterns through modelling and optimisation in order to find the best parameters for a given model. The patterns that we are trying to understand are the movement patterns of road users.”

In other words, machine learning is a process of finding patterns. In this example, a mathematical model for road user intention is built and then the model is fine-tuned by optimising its parameters until it represents the data in question the best.

Building models

Sensor positions and approximate fields of viewSensor positions and approximate fields of view In his project, Akshay used data derived from sensors previously installed in B-huset at Campus Valla to track the autonomous shuttle and other road users. The data he received (point positions) was collected every 0.2 second over a week in September 2021 during another project at the campus. 

Using sets of data from the sensors called training data, Akshay built three different prediction models. The accuracy of each model was then evaluated against other data sets from the sensors, the validation data.

“I started with a simple dynamics model based on Newtons laws of motion and once I understood what the data looked like, I went into a more complex machine learning model. It was a lot of learning for me as well. I started from a very basic model and then I moved on to more advanced models and through this process I learned more.”

What did the most advanced model learn?

For example, the model learned to predict when a person was going to self-correct their course. In some situations, collision avoidance between two persons could also be predicted.

“There were some interesting results. For example, the model learned to pay attention to someone who is close rather than someone who is farther away. That was really interesting because this is something that is very basic to us humans to know that, to pay attention to people who are close to you and then adjust your trajectory based on that. We found significant evidence that the model considers distance to other people in estimating its own [a person’s] trajectory.”

Akshay believes he stopped his research at the time when things were really getting interesting.

“This is a very fast developing field and this model did learn some basic patterns, but I feel that there is still a long way to go.”

Finally, Akshay wishes to emphasise the contributions to the project, from technical ideas to enthusiastic discussions, of his supervisors i.e., Per Sidén (IDA, LiU) and Yury Tarakanov (Viscando).

“My supervisors were incredibly helpful”, says Akshay, who today works for an autonomous driving company at Linköping Science Park.

Akshay Gurudath will receive SEK 10,000 at the award ceremony on the 29th of November at Campus Valla.

More about the award Show/Hide content

Contact Show/Hide content

Master's programme in statistics and machine learning at LiU Show/Hide content

LiU's artificial intelligence (AI) initiative Show/Hide content

Latest news from LiU Show/Hide content