Long-Range Gesture Recognition Using Millimeter Wave Radar

Abstract

Millimeter wave (mmWave) based gesture recognition technology provides a good human computer interaction (HCI) experience. Prior works focus on the close-range gesture recognition, but fall short in range extension, i.e., they are unable to recognize gestures more than one meter away from considerable noise motions. In this paper, we design a long-range gesture recognition model which utilizes a novel data processing method and a customized artificial Convolutional Neural Network (CNN). Firstly, we break down gestures into multiple reflection points and extract their spatial-temporal features which depict gesture details. Secondly, we design a CNN to learn changing patterns of extracted features respectively and output the recognition result. We thoroughly evaluate our proposed system by implementing on a commodity mmWave radar. Besides, we also provide more extensive assessments to demonstrate that the proposed system is practical in several real-world scenarios.

Millimeter Wave Sensing: A Review of Application Pipelines and Building Blocks

“…Several CFAR types have been considered. One dimensional types include cell averaging [93], [114], [121], [123], [124], cell averaging smallest of [55], [71], [96], [114], cell averaging greatest of [71], clutter map [83], [84], and ordered statistics [92]. A two dimensional custom CFAR type is considered in [131].…”

“…The feature extraction phase is, however, not a required step. Recent deep learning techniques [47], [103], [104], [123], [124], [149], [194] and several modeling algorithms [56], [159], [184] perform well on raw data samples and, therefore, do not require the feature extraction phase. The border between feature extraction and analytical modeling has become blurry in recent papers on millimeter wave sensing applications.…”

“…The pipelines therefore rely on experience gathered with images in the vision and object detection domains. In addition, a few approaches have tried using a temporal CNN with 1D profile [158] data and a feature distributed CNN with a single point cloud frame [124] as input.…”

“…Not adapting the learning rate, i.e., a hyperparameter that controls how much parameter update is used to change the model parameters, during training cycles can result in long execution time for model training or training instability. To combat these problems, adaptive learning rate techniques such as Adaptive Moment (ADAM) estimation [46], [47], [53], [63], [76], [92], [103], [123], [124], [126]- [128], [130], [158] or manual learning rate decay [125], [174] can be employed. c) Cross Validation: Cross validation refers to training a model and evaluating model performance multiple times on a variety of different data or feature set splits.…”

“…Training instability examples include the cost or reward function prematurely becoming constant, a cost function that suddenly starts to increase with continuous acceleration, etc. Batch normalization [16], [63], [92], [123], [124], [128], [138], [194], initializing the model parameters prior to training with Xavier initialization [126]- [128], [194], making changes to existing model types with for example the ResNet layer [92], [123] or LeakyRELU activation function [48], [53], [77], [138] and parameter update averaging over multiple training cycles with momentum [120], [125], [181], [195] have been proposed to solve the problem of training instability.…”