Vision system is very important for mobile robots to sense surrounding environment and navigate intelligently, superior to the navigation system without vision. Although great progresses have been achieved, there are lots of problems needing to be resolved on robot vision. The efficiency of the vision system and real-time performance of algorithm is the main bottleneck against robot vision system working in practice. This thesis is focus on the study of parallel image processing algorithms and their hardware realization. The novel work and contribution of this thesis can be summarized as follows: 1. Both the mobile robot vision system and the vision processing unit are retrospected and prospected. Based on the technologies of DSP and FPGA, an embedded vision system of high performance and low power consumption is developed, which has the architecture of modularization and can provide various communication interface. 2. A new hardware architecture for Gaussian convolution is presented. Making use of symmetry and seperability of Gaussian convolution kernel, we adopt two 1D Gaussian convolution of variable length instead of 2D structure. This architecture can handle the processing of the image border and improve hardware resource utilization . 3. The hardware architectures of an edge detection algorithm and a Hough based line detection algorithm are presented. In this architecture, the image can be processed at the same time when the image is captured. Thus the processing time is reduce dramatically compare with the serial processing method. 4. The object detection algorithm based on Haar feature and AdaBoost algorithm is analyzed thoroughly. A hardware realization of an object detection architecture is then presented. In this architecure, systolic array structure is used in order to compute the classifier in parallel.
修改评论