A 58.6 mW 30 Frames/s Real-Time Programmable Multiobject Detection Accelerator With Deformable Parts Models on Full HD $1920\times 1080$ Videos

This paper presents a programmable, energy-efficient, and real-time object detection hardware accelerator for low power and high throughput applications using deformable parts models, with <inline-formula> <tex-math notation="LaTeX">$2\times $ </tex-math></inline-formula> higher detection accuracy than traditional rigid body models. Three methods are used to address the high computational complexity of eight deformable parts detection: classification pruning for <inline-formula> <tex-math notation="LaTeX">$33\times $ </tex-math></inline-formula> fewer part classification, vector quantization for <inline-formula> <tex-math notation="LaTeX">$15\times $ </tex-math></inline-formula> memory size reduction, and feature basis projection for <inline-formula> <tex-math notation="LaTeX">$2\times $ </tex-math></inline-formula> reduction in the cost of each classification. The chip was fabricated in a 65 nm CMOS technology, and can process full high definition <inline-formula> <tex-math notation="LaTeX">$1920\times 1080$ </tex-math></inline-formula> videos at 60 frames/s without any OFF-chip storage. The chip has two programmable classification engines (CEs) for multiobject detection. At 30 frames/s, the chip consumes only 58.6 mW (0.94 nJ/pixel, 1168 GOPS/W). At a higher throughput of 60 frames/s, the CEs can be time multiplexed to detect even more than two object classes. This proposed accelerator enables object detection to be as energy-efficient as video compression, which is found in most cameras today.