RAODAT: An Energy-Efficient Reconfigurable AI-based Object Detection and Tracking Processor with Online Learning

Abstract

Smart robots (e.g. drones) for object detection & tracking demand for embedded intelligent processors. Neural network (NN) processors have been designed to accelerate NN for pattern recognition [1] [2]. However, these designs lack special processing engines for object detection & tracking such as bounding box (bbox) calculation and selection. Also, their architectures are designed for general AI tasks resulting in redundancy/inefficiency in performing object detection & tracking. An object detection processor has been proposed previously [3], but it only supports specific detection NN and does not support object tracking. Object tracking processors have also been proposed [4] [5], but these designs do not support object detection and thus cannot be used for object search. This paper presents RAODAT, which to the best of our knowledge is the first reconfigurable AI-based object detection and tracking processor with online learning. It has three key features: 1) An object detection & tracking architecture with reconfigurable NN and detection/tracking engines for programmable object detection & tracking tasks, 2) An object learning architecture with shared NN inference/learning engine and automatic label generation engine to support object tracking with online learning, 3) Layer- & stride-aware computing techniques to improve the NN computation efficiency.