Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Elmaghraby, Adel

Committee Co-Chair (if applicable)

Imam, Ibrahim

Committee Member

Imam, Ibrahim

Committee Member

Zhang, Hui

Committee Member

Losavio, Michael

Committee Member

Sierra-Sosa, Daniel

Author's Keywords

pedestrian detection; tracking; re-identification; virtual outlier synthetic: in distribution; out of distribution


Multiple Object Tracking (MOT) is the problem that involves following the trajectory of multiple objects in a sequence, generally a video. Pedestrians are among the most interesting subjects to track and recognize for many purposes such as surveillance, and safety. In the recent years, Unmanned Aerial Vehicles (UAV’s) have been viewed as a viable option for monitoring public areas, as they provide a low-cost method of data collection while covering large and difficult-to-reach areas. In this thesis, we present an online pedestrian tracking and re-identification from aerial devices framework. This framework is based on learning a compact directional statistic distribution (von-Mises-Fisher distribution) for each person ID using a deep convolutional neural network. The distribution characteristics are trained to be invariant to clothes appearances and to transformations. In real world scenarios, during deployment, new pedestrian and objects can appear in the scene and the model should detect them as Out Of Distribution (OOD). Thus, our frameworks also includes an OOD detection adopted from [16] called Virtual Outlier Synthetic (VOS), that detects OOD based on synthesising virtual outlier in the embedding space in an online manner. To validate, analyze and compare our approach, we use a large real benchmark data that contain detection tracking and identity annotations. These targets are captured at different viewing angles, different places, and different times by a ”DJI Phantom 4” drone. We validate the effectiveness of the proposed framework by evaluating their detection, tracking and long term identification performance as well as classification performance between In Distribution (ID) and OOD. We show that the the proposed methods in the framework can learn models that achieve their objectives.