The ASSET-2 system has been implemented at DRA Chertsey to run live on the ROVA PowerPC-based image processing system (see Figure 8). Feature detection is
Figure 8: The real-time implementation of ASSET-2 on the ROVA
PowerPC-based image processing system.
performed by custom hardware supplied by Roke Manor Research Limited, which provides a list of features at full frame rate (25 Hz). This custom hardware comprises four VME boards with Datacube Maxbus digital image bus connections on the front panels. A Datacube Digimax board digitizes each image and stores this digital image in a Datacube Framestore. The digital image is then sent from the Framestore down the Maxbus link to the Roke Manor boards which carry out the different stages of the Harris corner finding algorithm [15]. A Parsytec Transputer link to VME interface board takes the resulting list of image corners from VME space and sends it down a Transputer link to the first of two Parsytec Transputer/PowerPC processing boards, straight into PowerPC memory.
The Parsytec PowerPC boards run PowerPC 601 processors at 80 MHz; two such boards are used here for real-time running of ASSET-2. The feature tracking stage of ASSET-2 runs on the first PowerPC board at full frame rate. This is possible due to the efficient design of all algorithms which make up ASSET-2. The resulting vector flow field is sent down a Transputer link to the second PowerPC board, where the remainder of ASSET-2 is run, also at full frame rate.
Results are graphically overlaid onto the images stored in the Datacube Framestore using the Parsytec VME interface board again. Images are not stored for longer than one frame period in this double-buffered Framestore, so that graphics are not overlaid on exactly the image which they correspond to. However, the incorrect positioning is barely visible by eye, and errors are reduced by ``predicting ahead'' the movement of all graphics, as image velocities are available to make this possible.