Toshiba announces the TMPV7608XBG, an image recognition processor that provides recognition and detection of vehicles and pedestrians at night. The new processor is capable of 1.9 Tera operations per second (TOPS) and integrates new feature descriptors that make use of color-based image information.
The new processor is said to realize a nighttime pedestrian detection as reliable as a daytime detection available with conventional vision systems. Toshiba’s original Enhanced CoHOG (Co-occurrence Histograms of Oriented Gradients) accelerators combine luminance-based CoHOG feature descriptors with color-based feature descriptors obtained using a newly developed technique. This enhancement leads to an improvement in the recognition accuracy, especially at nighttime and at scenes with less luminance differences between objects and the background:
The TMPV7608XBG incorporates a Structure from Motion (SfM) accelerator that allows detection of general stationary obstacles such as fallen objects and landslides. The SfM accelerator provides three-dimensional (3D) estimates of the distance to, and the height and width of, the stationary obstacles, based on a sequence of images from a monocular camera. This accelerator makes it possible to detect any stationary obstacles without a learning curve, as well as moving objects (using motion analysis) and a particular class of objects such as pedestrians and vehicles (using pattern recognition).
The TMPV7608XBG is able to simultaneously perform Traffic Light Recognition (TLR), Traffic Sign Recognition (TSR), Lane Departure Warning (LDW) and Lane Keeping Assist (LKA), Vehicle and Pedestrian Collision Warning and Collision Avoidance, High-Beam Assistance, and General Obstacle Collision Warning. The devices interfaces with up to 8 cameras, while consuming 3.37W of power:
The sample shipment started in January 2015.