Master Thesis - Tracking Multiple Extended Objects with Deep Learning Detections
Master Thesis Opportunities – Gothenburg, Sweden
Autonomous driving is currently a very hot topic in research and development. The vehicle industry in the Gothenburg area, with actors such as Zenuity and Chalmers, is one of the global front runners in the race towards autonomous vehicles that are capable of not just highway driving, but also inner city urban driving. One of the enabling technologies for reliable and safe urban driving is high performance environment perception. Part of the environment perception is the autonomous vehicle’s capability to track other moving objects, such as cars, pedestrians and bicyclists. This is the so called multi object tracking (MOT) problem. Typically, an MOT algorithm builds upon object detection together with a framework for handling multiple objects, where the number of objects is not just unknown, but also time-varying.
Object detection is a problem where data is processed such that objects of interest are identified in the data. In the context of autonomous cars, the data can be images from a camera mounted on the vehicle, and the objects of interest can be other road users, such as cars, pedestrians, and bicyclists. In the past few years, image detection has been revolutionized by the advances in Deep Learning (DL), leading to object detection algorithms that can outperform humans.
In MOT research, so called Random Finite Set (RFS) methods have been a strong trend in the past decade, and have recently lead to the development of so called MOT conjugate priors. The RFS based MOT conjugate priors require models of the sensors. Previous work has shown that DL can be an efficient tool for constructing functions that pre-process the sensor data into detections that are input in a MOT algorithm. However, prior work did not fully take the car extent (its size and shape) into account, something that lead to a degradation of tracking performance in certain scenarios.
We are looking for two Master Thesis students with good grades in courses like Sensor fusion and Deep Learning.
This project lies on an interesting intersection between DL and MOT. It will focus on the tracking of cars, using image data that, using DL, has been pre-processed into detections. The detections will be bounding boxes (BBs) that indicate where in the image the car is; and for each BB estimates or the range to, and the heading of, the detected car. It is possible to extended this to also include information about the appearance of the detected car.
Regarding the MOT, an important task will be to include the car extent in the modelling, such that it can be estimated in addition to the position and the motion properties (velocity, heading, etc). A reasonable simplifying assumption here is that the tracked cars have a rectangular shape that is aligned with the heading. Including the target extent in the modelling will require the development of suitable measurement models, as well as the integration of those models into an MOT algorithm. Suitable motion models are coordinated turn models, or simplified bicycle models.
The tracking algorithm needs to be both computationally efficient and have accurate performance; recent advances in RFS based MOT have shown that so called Poisson Multi Bernoullli filters have both qualities. Such tracking frameworks facilitate the estimation of both the unknown number of cars, as well as the estimation of the state of each car.
The work will be evaluated using publicly available benchmark data. If time permits, the project can be extended to include other object types, such as pedestrians and bicyclist.
Further information and contacts
Final application date: December 15th 2017. Please send in individual applications, including grade transcripts. If you wish to partner with someone, simply state that in your cover letter.
Planned start: Beginning of 2018, with some flexibility.
Duration: 30 ECTS
For questions regarding the project, please Contact Louise Bichler, firstname.lastname@example.org.