ZYNQ flexible platform for object recognition & tracking

This paper presents the use of ZYNQ-7000 All Programmable SoC for flexible object recognition applications targeted for indoor mapping and localization. The architecture of the system is designed to provide the necessary infrastructure to support hardware software partitioning. Vivado HLS OpenCV libraries are used to synthesize the hardware for accelerating parts of the algorithms. Idea of identifying door handles serves as an opportunity to find paths for further navigation. Therefore, door handles and knobs served as target objects to be located in an image scene of the university corridor as a sample use case. The algorithm which performs object detection is implemented as a image data processing chain consisting of Sobel filter unit, Hough Line Transform unit and Oriented FAST and Rotated BRIEF unit. The Hough Line Transform extracts the lines from image scene to further locate the region of interest, while Oriented FAST and Rotated BRIEF extracts the image feature. Partitioning of algorithm is performed iteratively by moving the computation intensive tasks to the Programmable Logic and executing the rest of the algorithm on Processing System to achieve a balance between available hardware resource and acceptable frame rates. The resulting system positively identifies the object of interest at acceptable frame rates and with optimal system resource utilization.