Recently, object recognition has been widely used for various applications such as robot vision system, autonomous vehicle control, surveillance camera, and image search engine like Goggle developed by Google. The object recognition applications are characterized by complex and data-intensive computations, which make it difficult to achieve real-time performance even on today’s state-of-the-art processors. Especially, real-time performance and low-power consumption are important requirements for embedded systems. In addition, programmability is also considered to deal with various recognition targets and algorithms. uch requirements of the mobile robot vision system motivate our research to design real-time object recognition processors, BONE-V series, with low-power consumption.
  To achieve real time object recognition, four different researches are carried out; Algorithm, Architecture, VLSI implementation, Humanistic intelligence.

  For algorithm research, we develop new object recognition algorithm to retain higher accuracy as well as lower computation cost than conventional works. It exploits visual attention process of human brain so as to increase recognition accuracy and speed. Based on the new algorithm, a heterogeneous many-core processor is designed with the visual attention engine adopting massively parallel architecture for high throughput. After that, power and task management techniques, such as DVFS, power gating and dynamic resource management, are applied to the VLSI implementation of the processor. To realize humanistic intelligence on our SoC, we employ the way of human acts, think, and learn which overcome limitations of A.I into our research. It consists of the machine learning algorithm, object prediction and inference, and thread optimization. These studies can be used for humanoid robot, autonomous vehicle control, and vision-based guided-missile platform.