2023IOTJ
1.Abstract
person identification using point-cloud from mmWave radar.
2.Introduction
Visual sensor is vulnerable to illumination condition, and it is unsuitable for spaces with privacy requiremets.
Wearable device is inconvenient and limited by application occasion,
Micro-Doppler signature is deficient in spatial information.
3D point-cloud with rich spatial informtion and profile information gets popular.
Motivation:
Inspired by the promising results in [29], we attempt to control feature distribution by actively learning and optimizing the unknown distribution of target behavior through network construction. Furthermore, due to the significant impact of feature representation on identification performance, we also consider enhancing the feature representation of the raw point-cloud.
PIpeline:
At first, an improved filtering method by dual-stage density-based spatial clustering of applications with noise (DST-DBSCAN) for point-cloud captured by millimeter-wave (mmWave) radar is raised for clearing noise and ir-relevant points. Then, these filtered points are inputted into PointNet++ for feature extraction and recognition. Besides, adversarial network contributes to optimize feature distribution by improving feature capturing capability of PointNet++. At last, losses from both PointNet++ and adversarial network are calculated and optimized for updating network parameters.
Contributions:
- In DST-DBSCAN, a self-adapted threshold and DBSCAN are applied to leach noisy and uncorrelated points, and the accumulation operation alleviates the sparsity and dispersion of point-cloud.
- AdPNet is designed to optimize distribution of features extracted by PointNet++, which encourages the characteristics acquired by the feature extractor of PointNet++ to be more similar on the same person but distinct between different people.
3.Methodology
3.1 preprocessing of point-cloud data
- remove the frame containning less points with threshold:
- the filtered point sets:
- the accumulation progress within T frames:
- eliminate the irrelevant points in the accumulated frame with DBSCAN:
- normalize the point-cloud by subtracting its center:
3.2 feature extraction and classifier training
PointNet++ is selected as the feature extractor and classifier: The sampling layer selects aset of points from the point-cloud, which are defined as the centroid of local area. Then, the grouping layer constructs a local region set by looking for adjacent points of the centroid. After that, the PointNet layer encodes a local region set as a feature vector.
the cross entropy is applied,
centralized constraint is proposed for tightening feature distribution
regularization term is added to training loss to alleviate the overfitting phenomenon and
improve network performance
3.3 AdPNet and Adversarial Training
AdPNet is resposible for influening the front module of pointnet++, and it contains data preparation module and the adversarial network
data preparation module conbinate the feature and assign label 1 for features from the same person and 0 for different. The labels are used to train the adversarial network
4.Experiment
4.1 dataset
10 participants
an open and empty area is selected for data collection.
39412 raw data
4.2 network
4.3 ablation study
4.4 compare with different models
词汇积累:
contour 轮廓
gait 步态
vulnerable to illumination condition 易受光照条件影响
Nevertheless 尽管如此
deficient 不足的
profile 轮廓
Nonetheless 尽管如此
susceptible to noise 易受噪音影响
intuitive representation 直观表示
leach noisy 过滤噪声
alleviates the sparsity and dispersion 减轻稀疏性和分散性
substantiate the effectiveness 证实有效性
autoregressive generation model 自回归生成模型
utilize device with fancy price 利用价格昂贵的设备
under most of circumstances 在大多数情况下
slight contribution 轻微贡献
eliminate edge irrelevant points 消除边缘无关点
scattered points 分散的点
alleviate the overfitting phenomenon 缓解过度拟合现象
exploited 利用
deployed 部署
句子积累:
it is believed that different motions can also be remapped into codes with different distributions, and the homologous property can be found in point-cloud.