dc.contributor.authorLee, Joseph Yuan Sheng
dc.date.accessioned2018-09-24T07:35:11Z
dc.date.available2018-09-24T07:35:11Z
dc.date.issued2018
dc.identifier.urihttp://hdl.handle.net/10356/76041
dc.description.abstractThis thesis describes a multiple object tracker algorithm for autonomous vehicles based on visual data. The algorithm in this thesis is based on the MDP tracker [1], with extensions to use SiamFC [2] to generate a lightweight encoding used as features for association and tracking. The transitions between tracked and lost states are formalized as a Markov Decision Process. One main improvement was to use association before attempting to track, reducing tracking cost if associations are successful. Further enhancement includes the use of GPU processing. When an association fails, an attempt is made to track the object by searching a small section of the image. The search algorithm utilizes the same encoding method, which keeps the computational cost low. The usage of reinforcement learning and Support Vector Machine (SVM) in the original MDP tracker was replaced with batch training with Feedforward Neural Network (FNN). The tracker algorithm described in this thesis demonstrates an average update rate that is above real-time, while maintaining high performance benchmark scores.en_US
dc.format.extent64 p.en_US
dc.language.isoenen_US
dc.subjectDRNTU::Engineering::Electrical and electronic engineeringen_US
dc.titleObject tracking for autonomous vehiclesen_US
dc.typeThesis
dc.contributor.supervisorJustin Dauwels (EEE)en_US
dc.contributor.schoolSchool of Electrical and Electronic Engineeringen_US
dc.description.degreeMaster of Science (Computer Control and Automation)en_US


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record