As various applications of object recognition has been advanced, much higher performance vision pro-cessing technology is required. Object matching process which consists of feature matching and feature cluster-ing process has been the bottleneck of the whole object recognition process because of its massive memory bandwidth that occurs during external database access. Locality Sensitive Hashing (LSH) algorithm has been frequently used for feature matching process due to its high feature matching accuracy. Previous works based on LSH achieved high accuracy but they cannot meet the Real-time requirement under Full HD resolution and 60fps frame rate.
To compensate for speed, Vocabulary Tree (VT) is proposed and widely used due to its ability to remove external DB access and facilitate fast speed. Nevertheless, VT has chronic problem that its object matching accuracy is rapidly degraded as the number and size of the database increases. To compensate for low accuracy, Vocabulary Forest (VF) algorithm which is composed of 4 differently learnt VTs and a combination logic is proposed in this paper.
The proposed VF is implemented in 65nm CMOS Logic technology operating at 250MHz under 1.2V supply voltage. The VF processor with 190kByte SRAM in total occupies 1.360mm x 1.692mm and consumes 33.8mW in average. Also, the proposed VF processor achieved 95.7% matching accuracy and 2.07 M-vec/s throughput.