Human Activity Recognition from Point Clouds Generated through a Millimeter-wave Radar

An extension work of https://github.com/nesl/RadHAR, the previous published work can be found in https://doi.org/10.1145/3349624.3356768.

This extended work is mainly contributed by Qiong Hu, supported by Akash Deep Singh and Ziqi Wang. The report of this work can be found in the cover letter.

• Objective:
  • Filter out noise signals from Radar point clouds
  • Retrieve and generate human figures
  • Recognize and classify human motions
• Methods:
  • Extract the characteristics of data distribution individually on three dimensions
  • Use statistic functions to fit the target signal area in the data distribution histogram
  • Filter out noise signals
  • Use shift-and-add to obtain target figure information
  • Algorithm can be found in filter.py
• Results/Applications:
  • The filtering algorithm can effectively filter out environmental noises, estimate the human figure center and generate human cuboid representation every two seconds
  • Using the center points from all the datas can generate human motion trajectory
  • Using the cuboid representation can estimate human figure size and skeleton representation
  • The filtering algorithm is generated from one single dataset from one walking experiment, and is applicable for other walking experiments and more motions without having pre-knowledge or changing parameters
  • Resulting images can be found in img/
  • Resulting datas of trajectory estimation and figure size estimation can be found in results.yaml
• Future work:
  • About Methods:
    • Fitting function for data distribution in Z dimension may converge, which may be replaced by a learning algorithm
    • The shift-and-add function still need debug about which point in every frame to align to the center from 60 frames
    • The figure skeleton is based on fixed human body proportion for now, which may be replaced by information extracted from the filtered data points
  • About Applications:
    • The figure size estimation is not accurate enough yet
    • When the algorithm is applied to unfamiliar dataset, there would still be several noises that are not completely filtered out
    • The final objective would be to extract accurate human posture and recognize/classify human motions