基本介绍
基本信息
- 题目
Gesture-Radar: A Dual Doppler Radar Based System for Robust Recognition and Quantitative Profiling of Human Gestures - 作者
Zhu Wang , Member, IEEE, Zhiwen Yu , Senior Member, IEEE, Xinye Lou, Bin Guo , Senior Member, IEEE, and Liming Chen , Senior Member, IEEE
研究问题
subtle arm gestures with less positioning or environmental dependence
主要insight
这一组手势集合较为简单,且相互之间有比较大的差异,所以容易找到一些定性或者定量的特征进行区分。例如提到的根据运动速度的方向可以分为明显靠近的手势,明显远离的手势和不确定的。特征越多,可区分的集合就越多,所以文章的主体就是在找尽量多的特征,从多个维度区分这些手势。另外,可以从KNN没解决的那部分问题入手深挖一下,在获得了高精度的信息以后,进一步做数据分析应该会获得更有用的信息。
主要贡献
- Robust gesture recognition using two radar sensors. Gesture-radar uses Doppler radar sensors, which are originally used for speed measurement, to capture and identify human gesture types and other details. In particular, to overcome shortcomings of the single radar-based system (e.g., signal blind area and environmental dependence), we leverage two vertically placed radars to collect complementary sensing data of gestures (i.e., both radial and tangential components) and propose corresponding methods to eliminate environment interference. 联合两对设备可以算贡献吗?消除环境干扰可以算一个,但是从文中貌似没有看到对环境干扰的专门处理
- Accurate and quantitative profiling of gesture details based on multi-layer analysis. In order to recognize gesture types accurately and efficiently, gesture-radar adopts a two-stage classification approach including a phase discrimination and a signal matching method. Meanwhile, to get more details about the gesture, we propose an efficient method to quantitatively estimate the angle or range of gestures by exploring and fusing the time-frequency spectrograms of two radar sensors.因为雷达本身对动作细节刻画就很好,所以如果能在此基础上拿到比较好刻画动作的信息确实是很好的
实现
- 两个24GHz多普勒雷达:24 GHz K-LC2 short range radars with the ST-100 radar development board
- The radar sensors outputs dual channel signals called I and Q
- sampling rate 44 100 Hz
处理过程(对照下图)
第一阶段
大采样率、大带宽可以提供更准确的速度信息
第二阶段
In the second phase, gesture-radar extracts motion segments by performing a series of signal processing procedures, and each segment corresponds to one user gesture. Gesture-radar will first minimize the interference of environmental noise using a low-pass filter, and then uses a rule-based double-threshold comparison method to extract possible gesture segments. Finally, the discrete wavelet transformation method is adopted to compress the extracted gesture segments, so as to reduce the computation cost. As this part is similar to our previous work [39], we will not provide detailed descriptions in this article.
- 减少环境噪声干扰,使用的是低通滤波器;确实可以消除高斯噪声,但是对于 less positioning or environmental dependence应该没有作用
- uses a rule-based double-threshold comparison method to extract possible gesture segments——依据相位的变化大小和正负确定运动方向(靠近远离还是不确定),结合两对设备就可以初步判断出一些手势并分成不同的子集;但是手势segment的方式在???
- 离散小波变换进行数据降维并提取特征
第三阶段
手势方向
adopt a phase discrimination approach to calculate the direction of the gesture——通过I、Q的相位情况判断多普勒频移确定运动方向,但是这和第二阶段第二步的区别是?
手势识别
a restricted dynamic time warping (R-DTW) algorithm for waveform similarity calculation and a k-nearest neighbor (kNN) classifier for gesture recognition
R-DTW是利用动态规划的思想衡量两个波形的相似度,利用这一“距离”作为KNN的度量标准。
第四阶段
use a spectrum analysis method to obtain more details of arm movements. It includes the following two practical calculation methods: extremum ratio method and area ratio method. Gesture-radar can finally get the angle, range or the cycle of gestures.
这里想获得手势的角度、范围和周期信息,目的是用来处理非标准手势;其insight是通过两对设备的速度之比求角度。同时,因为多普勒雷达是对整个胳膊进行处理,所以难以分清各个部分的速度。为此,将时域的波形转化为频域的时频谱图。
- extremum ratio method:通过峰值之比求角度
- area ratio method:通过红色面积之比(去除offset,即低频低于20Hz部分)
理解浅陋,如有错误还望告知,不胜感谢!