摘要
As the fundamental technology of autonomous vehicles and high-speed tracking, high-speed vision always suffers from the bottlenecks of on-chip bandwidth and storage due to the resource constraints. To improve the resource efficiency, we propose a hardware-efficient image compression circuit based on the vector quantization for a high-speed image sensor. In this circuit, a self-organizing map is implemented for the on-chip learning of codebook to flexibly satisfy the requirements of different applications. To reduce the hardware resources, we present a reconfigurable complete-binary-adder-tree, where the arithmetic units are reused completely. In addition, a mechanism of partial vector-component storage is adapted to make the compression ratio adjustable. Finally, a parallel-elementary-stream design ensures a high processing speed. The proposed circuit has been implemented on the field-programmable gate araray and also applied in a high-speed object tracking system. The experimental results indicate that it achieves an encoding speed of 722 frames/s with 128 weight vectors when working at 79.8 MHz, and the worst tracking error caused by the proposed circuit is merely 9 pixels. These results evince that our proposed circuit can be completely integrated with a high-speed image sensor and used in high-speed vision systems.
