Related Work

Hand Tracking Applications

Hand Tracking App

Motion-based: IMUs

• pretraining
• error accumulation
• single point of instrumentation
• employs supervised learning

Vision-based: IR cameras/ laser

• fine-grained
• gesture model
• robust against positional drift
• privacy, power consumption, computation

RF-based: RF

• NLOS issues
• power requirements

Other

Microphone: SaveFace

• low-cost
• heavy training
• binary classification

Ultra Sound:

• finger movements

Magnetic Tracking

Sensing Magnetic Induction

Sensing the current induced on a coil by a magnet

• high tracking accuracy
• long sensing range
• requires an energy-intensive magnetic field generator
• large induction coil

Tracking Electromagnetic Field

Each electromagnetic generates an oscillating magnetic field at specific frequencies.

This method is also using magnetometers

• richer channel info
• requires wearing powered electromagnetics

Tracking Passive Magnet

LM-based method: array of magnetometers (16)

• no orientation
• machine-learning-based
• pre-defined pose

Algorithm

Tracking Algorithm

$\vec{B} = \frac{\mu_0}{4 \pi} \times ( \frac{3( \vec{m} \cdot \vec{r} )\vec{r}_{ij}}{|\vec{r}|^5} - \frac{\vec{m}}{|\vec{r}|^3})$

对于特定的偶极子就是求6个参数$x,y,z,m,\theta,\phi$ 每个magnometer的观察量都可以被认为是背景磁场和所有无源磁体的线性组合

$\vec{B_i} = G + \sum_{j=1}^{M} \frac{\mu_0}{4 \pi} \times ( \frac{3( \vec{m_j} \cdot \vec{r_{ij}} ) \vec{r_{ij}}} { | \vec{r_{ij}} |^5 } - \frac{ \vec{m_j} }{ |\vec{r_{ij}}|^3 })$

• 滑动窗口在输入滤波
• 卡尔曼滤波器在输出时进行轨迹的平滑和预测

Diminishing Magnets

模拟环境生成数据

SVM分类

(我们应该偏好 1or2 有高recall)

Hardware Design

• accuracy
• practical range
• low energy consumption and manufacturing cost

CAMAD

multi-layer -> 2 layer

computer aid vs proposed

PSO

Hardware Configuration

magnetometers calibration

Evaluation

implemented optimized LM in C++, then the algorithm is invoked by python.

Pilot Study

Using Leap Motion observation as ground truth.

框架变换

Overhead

diminishing vs ongoing

拓展

• 手语翻译
• 手语学习
• 康复训练
• 太空

瞎说的

• 需要手势的

• 乐器

• Channels(不同的磁矩大小？/ 初始对应)

• 手腕上距离磁铁距离不变为什么要diminish

• 为什么使用PSO？

• 距离越近偶极子的模型越不贴近实际磁场，距离远难以精准测量(线性叠加也要考虑各自磁体之间的影响)

• PCB板 -> 平板设计，难以设计成环

• 八边形环（纯纯的瞎说）

• ANN逆向求解(其实感觉不是很靠谱,无法保证精度，无法进行sensor拓展)