Master Thesis: Optimization of Multi Object Tracking Algorithms

Background - The Why

More than one million people die in traffic accidents every year. Advanced driver assistance systems (ADAS) provide functionality that assists the driver and warns and performs mitigating actions if dangerous conditions occur. A central part of these systems is the sensor fusion stack that combines the information from multiple data sources such as camera, radar and lidar into one common representation of the surrounding environment.

The number of active sensors on vehicles are today rapidly increasing which drives the development of more effective fusion algorithms. Such algorithms are often complex and put a high load on the real time computational processing units (CPUs) where they are allocated. Therefore, a challenge when developing the fusion algorithms is to balance functional performance with runtime performance. More complex algorithms with high functional performance could sometimes not be feasible to implement due to constraints on CPU load budgets. The selection of a correct fusion algorithm is therefore of high essence - to get as balanced performance as possible.

In an automotive setting there are many objects of interest. Some examples include pedestrians or vehicles, but also stationary objects such as guardrails and obstacles. An important area in perception is keeping track of these objects over time, typically achieved by using Multi Object Tracking (MOT) algorithms.

Who are we looking for?

Student(s) with a passion for traffic safety who are curious, enthusiastic, collaborative and forward leaning.

We prefer student(s) with a strong background in sensor fusion, machine learning, programming in MATLAB/Python/C++ and good mathematical background. Prior experience with control theory and modeling/simulation is meritorious.

Description of thesis work

The purpose of this thesis is to:

1. Research current state of the art algorithms in the field of object tracking. Classical or deep learning methods could be studied.
2. Investigate if the algorithms could be used for high-level object tracking and which properties would be needed from the sensors.
3. Implement, test and benchmark different algorithms on real sensor data. Performance indicators could for example be Object tracking performance and computational cost.

Thesis Level: Master

Language: English

Starting date: January 2023

Number of students: 1-2

Tutor

Daniel Johansson, Function Developer, +46 73 902 43 12

Felix Gustavsson, Function Developer, +46 73 902 61 52

Alexander Greger, Function Developer, +46 73 902 72 14

Leo Laine, Innovation and Research Strategist, +46 31 323 53 11