Get email delivery of the Cadence blog featured here
Single-digit semiconductor market growth pales in comparison to the double-digit growth projected for imaging and computer vision applications—an anticipated next driver of SoC growth. Advances in imaging/vision-processing algorithms and related IP are creating incredible market opportunities for semiconductor companies, noted Himanshu Sanghavi, a design engineering group director at Cadence, during his April 5 talk at CDNLive Silicon Valley’s Tensilica Track.
“Imaging refers to making an image more visually appealing,” said Sanghavi. “Computer vision deals with trying to extract more information out of that image.”
This revolution in visual systems stems from various factors, from the availability of inexpensive, fast, and high-resolution sensors to increases in compute density and processing power and the fast evolution of imaging algorithms. All of this has led to new application areas like drones and robots, advanced driver assistance systems (ADAS), security, and mobile systems for recognition and detection, Sanghavi explained.
Growing Sophistication of Computer Vision Algorithms
Computer vision algorithms have continued a steady upward trend in sophistication. Early on, there was feature detection and mapping. Then came algorithms that could estimate 3D structures from 2D image sequences. Soon, we had detection and recognition. Now, we’re seeing algorithms, using convolutional neural networks (CNN), deliver high recognition and prediction of “almost anything,” noted Sanghavi.
While cameras are collecting loads of visual data, much of the intelligence within this data lies untapped. Security applications provide a good example. “There’s a large number of cameras being deployed and capturing video 24/7,” said Sanghavi. “What you want to be able to do is run analytics on it so it tells you when there is something interesting happening in this video.”
The SoC architectures required to implement these multi-core imaging/vision systems will include a vast array of IP to support vision, audio, and fusion functions as well as a variety of interface protocols. Cadence’s Tensilica Vision P5 DSP is an example of IP that provides the performance and energy efficiency to support complex imaging and computer vision algorithms.
Tensilica Vision P5 block diagram
Uurmi Systems: Situational Awareness for Self-Driving Cars
In separate Tensilica Track sessions, two Vision P5 customers shared their experiences using the DSP with CDNLive attendees. Swarup Medasani, CTO and co-founder of Uurmi Systems, talked about his company’s semi-global block-matching (SGBM) algorithm for stereo-based range detection. Traditionally, such algorithms utilize correlation or block matching, which works fine for frontal, planar views of scenes. But autonomous vehicles need to see objects at multiple depths. Uurmi’s SGBM algorithm focuses on multiple directions to generate smoother, more dense disparity maps, which can enhance situational awareness for self-driving cars. Because the algorithm is complex, Uurmi broke it into smaller elements for processing, achieving 38fps on 240p and 11fps on 480p at a 600MHz clock rate on the Vision P5 DSP.
Sensory: Face Authentication
Dr. Paul Mullowney, senior research scientist at Sensory, discussed his company’s face-authentication algorithm, which he and his team are porting to the Tensilica IVP-EP DSP (the predecessor to the Vision P5 DSP). Mullowney noted that on the IVP-EP DSP, the algorithm demonstrated high accuracy and 12X faster runtime for face transform, feature computation, and logistic regression scoring compared with a 32-bit embedded processor. Next on tap, Mullowney plans to implement a robust pupil/face detection algorithm and port it to the Vision P5 DSP.
As Sanghavi noted, “Vision systems are becoming a key distinguishing factor for all of these systems.”
If you’re planning to attend CDNLive EMEA in Munich May 2nd-4th, be sure to make room in your schedule for the Tensilica and Automotive Tracks there.
Driving Automotive Vision Applications with the Right DSP
Q&A: Drones, Robots, and the New Tensilica Imaging/Vision DSP