A Real-time Scalable Object Detection System using Low-power HOG Accelerator VLSI

Abstract

As described in this paper, a real-time object detection system using a Histogram of Oriented Gradients (HOG) feature extraction accelerator VLSI is presented. The VLSI [1, 2] enables the system to achieve real-time performance and scalability for multiple object detection under limited power condition. The VLSI employs three techniques: a VLSI-oriented HOG algorithm with early classification in Support Vector Machine (SVM) classification, a dual-core architecture for parallel feature extraction, and a detection-window-size scalable architecture with a reconfigurable MAC array for processing objects of different shapes. The test chip was fabricated using 65 nm CMOS technology. The measurement result shows that the VLSI consumes 43 mW at 42.9 MHz and 1.1 V to process HDTV (1920?×?1080 pixels) at 30 frames per second (fps). A multiple object detection system and a multiple scale object detection system are presented to demonstrate the system flexibility and scalability realized by VLSI and applicability for versatile application of object detection. On the multiple object detection system, a real-time object detection for HDTV resolution video is achieved with 84 mW of power consumption on a task to detect 2 types of targets while keeping comparable detection accuracy as software-based system. On the multiple scale object detection system, a task to detect 5 scales of a target is accomplished using a single VLSI. The power consumption of the VLSI is estimated to 102 mW on the task.