Multi-object tracking coprocessor for multi-channel embedded DVR systems

Abstract

In this paper, the architecture of a video analytics coprocessor is proposed for multi-channel embedded digital video recorder (DVR) systems. A reference video analytics algorithm is proposed for multi-object tracking and is divided into independent processing steps based on data flow. Each step is designed in hardware or software considering its computational complexity and required system resources. Pixelwise processing requiring a large amount of computational resources, such as frame difference and background modeling, are designed as hardware with embedded direct memory access (DMA) controllers. A single-pass connected component labeling (CCL) is designed as a hardware targeting real-time processing of stream input. High-level tasks such as object filtering, frame-based control of hardware modules, and communication with an external host are designed with software on an embedded processor. Object tracking and event detection are designed with software on a host processor. Considering both the bandwidth required for frame processing and the bandwidth available by memory buses, the architecture of a 4-channel video analytics coprocessor is explored. It is finally implemented on a field-programmable gate arrays (FPGA) device, integrated into a conventional DVR system, and verified as to its functions and performance. It can provide video analysis functions to conventional DVR system-on-chip (SoC), and can lessen the cost of real-time video monitoring at remote monitoring centers.