021_EETE

EETE APRIL 2013

AUTOMOTIVE SAFETY not need additional frame memory on chip. As well as HDR, the TCM5114PL incorporates a number of other important functions such as auto luminance control, auto exposure, auto white balance, lens shading correction, auto flicker detection and correction and bad pixel correction. These minimize the need for external components and reduce requirements for additional software processing. The availability of digital and Fig. 2: Visconti3 block diagram. NTSC/PAL format analogue outputs simplifies integration with a wide variety of display panels. Variants with overlay functions and specific features to comply with the USA’s ‘back-over’ prevention regulations are under development. The need for higher resolutions is addressed by Toshiba’s 1080p full HD TCM5117PL CMOS image sensor, which combines a 2.7μm pixel size with colour noise reduction technology to ensure high sensitivity, better visibility and colour-rich images in low-luminescent environments. Image processing Once the image has been effectively captured in a range of light and speed conditions it must be processed in real time. This is where dedicated automotive image recognition processors are becoming important. As an example, figure 1 shows an automotive vision system based around the latest generation of Toshiba’s Visconti image recognition processor family, Visconti3. Built around a 32-bit multicore processor, Visconti3 has a video input interface supporting connectivity for up to four camera inputs and combines a multi-parallel media processor architecture with multiple image processing accelerators. A video output interface provides information to a dashboard or console LCD panel, while CAN connectivity supports direct connection to an automotive CAN bus. As the block diagram in figure 2 shows, the device supplied in a compact 27x27mm 516-BGA package also incorporates on-chip memory, a controller for external NOR Flash/SRAM and DDR2 SDRAM, a singlelane PCI Express interface, an 11-channel timer and SPI, UART and I2C interfaces. Image recognition algorithms and software in embedded image recognition processors must analyze large volumes of image data from the video source, frame-by-frame in real time, but with limited power consumption and memory resources. Devices without floating point calculation abilities can struggle with some algorithms, which is why Visconti3 uses a dual-core ARM Cortex-A9 MPCore design with a single/double-precision FPU (floating point unit) integrated into each core. The ‘de-facto standard’ ARM Cortex-A9 implementation works in conjunction with Toshiba’s proprietary high-performance 32-bit RISC CPU Media embedded processor (MeP). This configurable core gives designers the flexibility to customize processors at the design stage, including the ability to change processor configurations and add custom instructions to satisfy application requirements. At the heart of the Visconti device is the ‘image recognition engine’, which comprises four multi-core Media Processing Engines or MPEs. Each of these MPEs brings together a 32-bit RISC core MeP and MeP coprocessor suitable for multimedia processing, plus I- and D-cache and integrated RAM. The MPEs Fig. 3: Histogram of oriented gradients (HOG) detection. use a multi-grain Parallelism Architecture and VLIW (Very Long Instruction Word) technology in which multiple instructions including SIMD (Single Instruction Stream-Multiple Data Stream) instructions operate simultaneously. Image processing accelerators The six accelerators incorporated into the Visconti IC provide dedicated, high-speed processing of the key functions needed for image processing. The affine transformation accelerator, for example, performs functions such as re-sizing, lens distortion correction and transformations based on look-up-tables or formulas to eliminate distortion from images displayed to the driver. A matching accelerator performs stereo processing, tracking and optical flow functions, while two filter accelerators further optimize system performance through functions ranging from noise reduction and smoothing to edge and corner detection and colour space conversion. The histogram functionality provided by Visconti’s accelerators allow for the identification of specific objects and, in particular, pedestrians. Using a technique known as ‘Histogram of Oriented Gradients’ or HOG, the accelerators compare detected images against known characteristics of human body shapes and movement to distinguish pedestrians from other, inanimate objects such as waste bins or post boxes. The technique is based on the premise that edges associated with anatomical features such as shoulders, arms, legs and hips have gradients within certain upper and lower limits and orientation with respect to each other can be detected within specific regions of the anatomy. The system detects the presence of humans by building HOGs as shown in figure 3. While the standard HOG detection rate is limited, Toshiba’s R&D teams have further developed the HOG-based algorithm by improving the detection rate of pedestrians. The latest Visconti processor takes this a stage further with a hardware accelerator that combines CoHOG or Co-occurrence of Histograms of Oriented Gradients functionality with a linear support vector machine. This accelerator uses pairs of HOGs, which are more descriptive than single orientation HOG and require less complex processing than alternatives such as ‘shapelet’ (combination of edges) recognition. CoHOG is seen as the most efficient HOG method for the reliable detection of pedestrians. Power consumption Finally, it is worth noting that despite the high-performance and complex processing required for image-based ADAS implementations, it remains essential for developers to keep power budgets to an absolute minimum. The approach taken by Toshiba has been specifically designed to achieve this objective, resulting in a typical power consumption of 1.x W, depending on the degree of processing currently demanded by the various applications running in parallel. 20 Electronic Engineering Times Europe April 2013 www.electronics-eetimes.com


EETE APRIL 2013
To see the actual publication please follow the link above