Abstract
Hand gesture recognition (HGR) has become a prevalent human–computer interaction (HCI) approach on smart edge devices. The existing HGR systems suffer from either extensive power consumption and latency or low recognition rates under complex backgrounds. This article proposes an ultralow power real-time HGR system for the Internet-of- Things (IoT) applications through algorithm and hardware co-optimization. On the algorithm aspect, this work utilizes color- and depth-based hand segmentation for better background removal under complex backgrounds. In addition, the proposed bi-directional convolution-based feature extraction (FE) improves the rotation resistance, and the iteration-free feature clustering scheme reduces processing latency. On the hardware aspect, the adaptive activation of the gesture recognition core (GRC) reduces the dynamic power of the system by eliminating unnecessary signal toggling. Moreover, the fully pipelined FE engines and the systolic feature clustering processors further reduce the processing latency. In addition, the cluster combiner removes the redundant comparisons and improves the computing efficiency by 46%. The proposed HGR processor fabricated in a 65-nm CMOS technology that consumes the lowest power of 181 µW at 0.58 V . It can recognize nine static gestures and 20 dynamic gestures with an average accuracy of 98% and 99.1%, respectively. In addition, it can track the fingertip with an average accuracy of 2.4 pixels at a distance of 15–60 cm.