A feature vector matching in object recognition is the process of finding nearest neighbor database vec-tor for a given feature vector. Since it needs lots of required external bandwidth, it becomes the main bottleneck of real-time object recognition.
To reduce the required external bandwidth, the proposed feature vector matching processor utilizes spatio-temporal locality of nearest neighbor database vector. In video environment, the majority of the nearest neighbor vectors are commonly founded in previous frames at similar location. To support the spatio-temporal locality of nearest neighbor vector, a special cache for feature vector matching, namely, Spatio-Temporal Data-base Cache (STDB Cache) is newly proposed. In addition, to reduce matching error induced from the spatio-temporal locality method, mixed-mode neuro-fuzzy cache controller is proposed.
As a result, the proposed feature vector matching processor achieves 125,582 vec/s throughput and 95.1% matching accuracy, which are 2.02x and 1.32x higher than the state-of-the-art respectively. Therefore, the proposed feature vector matching processor achieves the most efficient throughput (vec/s o accuracy) enabling real-time object recognition for VGA 30fps video streams.