Multi-camera Multi-object Cross Tracking in Urban Environment.

Aim: This project aims to develop a machine learning based framework for multi-camera multi-object tracking (MCMOT) of vehicles and people in urban environment.

Background: Vision based object tracking has constantly attracted considerable research attention due to its significant commercial potential despite its technical challenges. Tracking is to analyse video sequences from single or multi- camera systems, while aiming to determine the location of targets over a series of frames. It plays an essential role in computer vision due to its capacity in several real-world applications, such as video surveillance, autonomous driving, human computer interaction, anomaly action detection, crowd behaviour analysis, traffic scene understanding, etc.

However, the presence of various photometric and geometric constraints has made visual tracking a challenging computer vision task. Nevertheless, with the availability of low-cost high-quality, high-frame-rate security cameras and ever-increasing computing power, the deployment of multi-camera surveillance systems for various applications has grown significantly.

While MCMOT system offers various advantages, it presents more technical challenges than single camera tracking, e.g. synchronisation of multi cameras, handling occlusion, object appearance changes etc. These challenges demand advanced computer vision techniques and sophisticated algorithms to ensure accurate and reliable object tracking across multiple cameras.

Therefore, this project aims to improve the performance of current algorithms and proposes a machine learning based framework for multi-camera multi-object tracking. To achieve the aim above, Objectives are listed as below:

O1: Single view object detection: implement current algorithm to accurately identify and categorize multiple objects within individual camera views, encompassing both stationary and dynamic objects.

O2: Intra-camera object tracking (single camera multi object tracking): Implement techniques for consistent and real-time tracking of multiple objects within a single camera's field of view, accounting for occlusions, abrupt motions, varying light conditions and different resolutions.

O3: Inter-camera object tracking (multi camera multi object tracking): Design a system to maintain object identity and tracking continuity across different camera fields of view, integrating spatial and temporal data to ensure seamless transition between cameras.