Detecting objects in videos and films is a challenging problem due to the motion of the subjects, the camera and the background. Additional problems are unreliable data caused by variations of illumination. Another challenge is to detect not fully visible but partially occluded objects as well. Moreover, if the targets of the detection are people, problems caused by the variation of appearance and clothing must be handled as well. In image processing, one bottleneck of algorithms has always been the lack of enough computational capacity. Hence real-time processing of video data is not possible in most of the cases using a general purpose CPU. To avoid this problem FPGAs can be used in many cases, since they take advantage of the parallel way of processing data. Designing systems which work in real-time is necessary in many areas of Computer Vision. For example: we take a vehicle in which there is an in-built system with a purpose of detecting pedestrians in front of and around the car. Both driver and pedestrians expect that the system works trustworthily and in real-time in any traffic situations in order to avoid accidents. The goal of the thesis is implementing a motion-feature-based pedestrian-detection algorithm on an FPGA-based platform. The motion must characterize human motion well while remaining resistant to camera and background motion.
In Chapter 1, a general, brief overview of Computer Vision research is described. In Chapter 2, an overview of state of art and the central ideas of the work are presented. In Chapter 3, the simulation of the Lukas and Kanade optical flow estimation algorithm is detailed, additionally simulations are presented to prove the performance of the design. Moreover, the Detection algorithm is described, including the details of the used pedestrian dataset and machine learning algorithm. In Chapter 4, the hardware setup of the FPGA implementation is discussed as well as the most important interfaces of the design. In Chapter 5, the FPGA implementation of the Lukas and Kanade single-level optical flow estimation algorithm as well as the motion HOG-based detection process is detailed. In Chapter 6, the achieved results of the work are presented and details of hardware resources of the FPGA implementation discussed. In Chapter 7, conclusions regarding the work are described as well as possible future improvements of the detection system.