Research in target pursuit using Unmanned Aerial Vehicle (UAV) has gained attention in recent years, this is primarily due to decrease in cost and increase in demand of small UAVs in many sectors. In computer vision, target pursuit is a complex problem as it involves the solving of many sub-problems which are typically concerned with the detection, tracking and following of the object of interest. At present, the majority of related existing methods are developed using computer simulation with the assumption of ideal environmental factors, while the remaining few practical methods are mainly developed to track and follow simple objects that contain monochromatic colours with very little texture variances. Current research in this topic is lacking of practical vision based approaches. Thus the aim of this research is to fill the gap by developing a real-time algorithm capable of following a person continuously given only a photo input. As this research considers the whole procedure as an autonomous system, therefore the drone is activated automatically upon receiving a photo of a person through Wi-Fi. This means that the whole system can be triggered by simply emailing a single photo from any device anywhere. This is done by first implementing image fetching to automatically connect to WIFI, download the image and decode it. Then, human detection is performed to extract the template from the upper body of the person, the intended target is acquired using both human detection and template matching. Finally, target pursuit is achieved by tracking the template continuously while sending the motion commands to the drone. In the target pursuit system, the detection is mainly accomplished using a proposed human detection method that is capable of detecting, extracting and segmenting the human body figure robustly from the background without prior training. This involves detecting face, head and shoulder separately, mainly using gradient maps. While the tracking is mainly accomplished using a proposed generic and non-learning template matching method, this involves combining intensity template matching with colour histogram model and employing a three-tier system for template management. A flight controller is also developed, it supports three types of controls: keyboard, mouse and text messages. Furthermore, the drone is programmed with three different modes: standby, sentry and search. To improve the detection and tracking of colour objects, this research has also proposed several colour related methods. One of them is a colour model for colour detection which consists of three colour components: hue, purity and brightness. Hue represents the colour angle, purity represents the colourfulness and brightness represents intensity. It can be represented in three different geometric shapes: sphere, hemisphere and cylinder, each of these shapes also contains two variations. Experimental results have shown that the target pursuit algorithm is capable of identifying and following the target person robustly given only a photo input. This can be evidenced by the live tracking and mapping of the intended targets with different clothing in both indoor and outdoor environments. Additionally, the various methods developed in this research could enhance the performance of practical vision based applications especially in detecting and tracking of objects.
Date of Award | 2018 |
---|
Original language | English |
---|
- drone aircraft
- computer vision
- automatic tracking
- mathematical models
- algorithms
- object tracking
Computer vision in target pursuit using a UAV
Su, F. (Author). 2018
Western Sydney University thesis: Doctoral thesis