Implementation of Canny edge detection on the WiCa SmartCam architecture

Abstract

There is a rapidly growing demand for cameras containing built-in intelligence for various purposes such as surveillance and identification. Face recognition is an important application for these cameras. Previous research has shown that faces can be well represented by parabola segments using an algorithm which fits parabola segments to edge pixels. Faces are then recognized using a technique which matches parabola segments based on distance and intensity. Considerable computational resources are required for the extraction of the parabola primitives, due to the need for Canny edge detection. This algorithm is well-suited for the new generation of massively parallel SmartCams though, if it is represented in a highly parallelized, low complexity representation adapted to the characteristics of these SmartCam architectures. This paper proposes such an implementation of the Canny edge detection algorithm for the Single Instruction Multiple Data (SIMD) Xetal IC3D processor, resulting in a real-time performance using a smart camera not bigger than a typical surveillance camera.