Automated Entry System using Multi-Object Tracking

Abstract: In this thesis we employ computer vision methods in order to extend and improve the functionality of automatic doors. This thesis is based around the implementation of a door entrance system which uses information from detected pedestrians to make qualified decisions regarding door activation. This can lead to a reduced amount of unnecessary openings which will reduce the energy consumption of buildings. It will also increase comfort for pedestrians passing through the door. A corner stone in the proposed system is multi-object tracking for which different methods are considered and evaluated. To provide input for the tracker a range of different detection methods are evaluated and used in the system. In order to tune and test this system a dataset consisting of realistic scenarios was collected and annotated. Results in the thesis show that we can estimate the walking direction of pedestrians well while the estimated speed is quite inaccurate. We also show that because of the good direction estimate one can employ static increases in prediction time to improve performance. Our tests show that YOLO, a modern object detector, is best at detecting pedestrians. We found that a tracker of relatively low complexity, Hungarian algorithm with Kalman filtering, receives both high scores and is quite robust to noise. It is concluded that this type of method can both extend and improve the automatic entry systems used today.