Pierre Paulin, R&D Director for Embedded Vision at Synopsys, presents the "Low-power Embedded Vision: A Face Tracker Case Study" tutorial at the May 2015 Embedded Vision Summit.
The ability to reliably detect and track individual objects or people has numerous applications, for example in the video-surveillance and home entertainment fields. While this has proven to be a challenging problem, recent years have brought higher performance solutions such as the Tracking-Learning-Detection (TLD) algorithm.
Pierre describes a face tracking application inspired by the TLD algorithm mapped onto an embedded vision solution. He explains the main concepts of the tracking algorithm and what was modified from TLD to better exploit the underlying hardware and stay within its memory and computational limits. He also describes the development flow from a functional to an optimized OpenVX capture. By comparing the optimized face tracking application running on the targeted embedded vision platform with a version running on an off-the-shelf 32-bit RISC processor, he demonstrates the power-performance-area gains that can be obtained using a solution tuned for computer vision.
