Per-pixel processor aims for real-time motion detection

Per-pixel processor aims for real-time motion detection

Mobile robots and self-guided vehicles often use image motion information to track targets and obtain depth information from scenes. Unfortunately, traditional motion algorithms running on Von-Neumann processing architectures are computationally intensive, preventing their use in real-time applications. Consequently, researchers developing image motion systems often turn to faster, more unconventional processing architectures.

One such architecture is the processor-per-pixel design, an approach recently proposed by Marc Tremblay and his colleagues at the Computer Vision and Systems

Laboratory of the Universit? Laval Sainte-Foy (Qu?bec, Canada). Fabricated in 1.5-mm CMOS and 0.8-mm BiCMOS, the low-resolution prototypes currently under development integrate a 50 x 50 smart-sensor array with integrated signal-processing capabilities.

OEach pixel integrates a photodetector, an analog signal-processing module, and a digital interface,O says Tremblay. OPixels are sensitive to temporal illumination changes produced by edges in motion. If a pixel detects an illumination change, it signals its position to an external digital module. Here, time stamps from a temporal reference are assigned to each sensor request. These time stamps are stored in local RAM and are later used to compute velocity vectors. The digital module also controls the sensor?s analog I/O signals and interfaces the system to a host computer through a serial link,O he adds.