Master by Research
University of Wollongong. School of Electrical, Computer and Telecommunications Engineering
Quelin, Mickael, Optical flow estimation using insect vision based parallel processing, Master by Research thesis, University of Wollongong. School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, 2011. https://ro.uow.edu.au/theses/3410
Computer vision systems are largely used in today’s industrial and technological worlds. The increasing complexity and precision of the visual information processed by modern methods allows for a wider range of applications in safety, surveillance, reconstruction and human-computer interaction. As the computing power available in hardware evolves, the trend in computer vision research is to create more precise reconstructions of the visual field by performing more extensive analysis and implementing more complex vision models. By contrast, in our approach, rather than trying to model the visual scene as accurately as possible, we wish to only access key information that is visual motion, at critical locations in the visual field. Our approach uses insect vision as a clue to design an intelligent motion detection system that can efficiently simplify the processing of visual information by splitting it into tasks that can be run in parallel.
Indeed insect vision has been studied extensively over the last century. The small brain size of these animals compared to the human one suggests a low complexity and fast processing neural system that is highly efficient for navigationin 3D environments. The elementary principle for motion detection in insects is well-known today and has been extensively studied. However, in order to deal effectively with 2D video signals that feature complex and changing environments in presence of noise, additional signal processing techniques need to be developed.
Therefore, the goal of this thesis is to develop a low complexity algorithm that is easily implementable on embedded hardware, an algorithm with fast processing and fixed latency. The developed system reads and processes the video signal derived from one camera input (monocular vision) to detect visual motion and convert it into an optical flow output format. Thus, our algorithm can be adaptedto many standard vision set ups that use motion information without the need for powerful computer hardware.
A recent digital model of the elementary motion detection principle, namely the template model, serves as a basis for our work. The main advantage of our approach is that it combines the power of parallel digital signal processing techniques with the elementary motion detection principles from insect vision. Some of the challenges addressed in this research are related to parallel velocity estimation, motion direction classification, robustness to noise and contrast, and simplicity and speed of processing. Moreover, being a low complexity system that deals with camera signals, the cost of a device that implements the developed algorithm would be attractive for many real-time applications. The developed algorithm could be used for further processing such as fast object detection, tracking and collision detection.