Energy-efficient context-aware real-time object recognition processor

Abstract

Nowadays object recognition has been widely adopted as a key component of various intelligent vision applications. For example, in the advanced driver assistance system, it enables the lane detection and the traffic sign recognition to improve the drivers safety or self-driving and, in the mobile phone, it enables the users to control their phone by just looking it through its camera without having to type commands. However, the object recognition requires a huge number of computations for the vision processing so it is difficult to achieve real-time performance on a mobile system with a general-purpose processor. In addition, the real-world environments make recognition even harder, where scenes usually contain objects varying in scale, illumination, distractors with similar shape, and occlusions on a cluttered background. We note the fact that human can recognizes thousands of objects without much difficulty, despite these variabilities. In this thesis, we aim to achieve highly reliable recognition accuracy by mimicking the human visual system. Human visual system not only makes use of local object information but also actively considers the global scene context. In the case of poor viewing quality such as clutter, noise, and blurred image, the scene context information play a major role in enhancing the reliability of the object recognition. In order to successfully achieve the human-like context-aware object recognition, we designed and developed (1) a multi-classifier recognition algorithm capable of recognizing the object and scene simultaneously, (2) a heterogeneous many-core architecture for real-time operation, (3) an analog-digital mixed-mode implementation for low-energy filtering, (4) a system board supporting fine-grain dynamic voltage and frequency scaling, and (5) a demonstration platform for the intelligent unmanned vehicle system. First, the proposed multi-classifier recognition algorithm utilizes Retinex pre-processing for illumination adapta...