A key task in image analytics is identifying regions-of-interest (ROI). ROI are areas in an image that are likely to contain objects of importance to the application. Once identified, ROI can be subsequently analyzed to detect the presence of these objects of importance. By focusing object recognition only on ROI, high system throughput can be maintained without loss of overall scene understanding functionality. This accelerator solution provides real-time region-of-interest detection capabilities for machine vision and video analytics applications. This accelerator would typically by the first stage of a visual object detection and classification system. This solution was demonstrated at OpenPOWER 2016 and SuperComputing 2016

Many machine vision algorithms use visual features for object detection. At a high-level, an object’s visual features represent a unique fingerprint that can be used to distinguish and identify the object in cluttered scenes. When searching an image for a particular object, the unique fingerprint is matched against all fingerprints in the image. SURF is a classical visual feature that has many applications in object detection, image registration, augmented reality and object tracking. Our FPGA-based accelerator computes SURF features for HD imagery while maintaining real-time frame rate of greater than 30 Frames Per Second