Embedded Human Activity Recognition Student Competition

Motivation and Background


Human Activity Recognition (HAR) has been a very active research field since it is the first step to monitor, diagnose, and cope with movement disorders [1]. The rise of smartphones and wearable devices enabled many new embedded system solutions on these platforms. However, access to a common set of experimental data has been a major obstacle that prevents breakthrough in this area. According to the International Technology Work Force on Parkinson’s Disease, two major technology barriers are the lack of compatible platforms and isolated research efforts [2]. We are organizing the embedded human activity recognition student competition to stimulate research and innovative solutions toward practical IoMT for health monitoring.

What is Provided to the Participants?


One of the major barriers to HAR research is access to labeled data sets. To this end, we developed an open-source wearable IoT platform for health and activity monitoring [3]. Using our prototype, we performed user studies with 22 users while they performed the following daily activities: Go downstairs, Go upstairs, Jump, Lay down, Sit, Stand, and Walk. The following dataset from these experiments will be provided to the participants:

  1. Raw 3-axis accelerometer data
  2. Raw stretch sensor data
  3. Activity labels for the raw data
  4. Example features for each labeled activity window
  5. A baseline reference HAR classifier [4] implemented in Matlab using this data

We provide precise time stamps and activity windows in contrast to coarse data sets collected by non-real-time platforms, such as smartphones. This is the only data set that has stretch sensors data. It is also one of the largest publicly available data sets.

What Will Be Submitted?


The goal of the contest is to design energy-efficient HAR classifiers. The participants start with the raw data with the provided windows. Then, the goal for the participants is to design an efficient feature set and a classifier. The designs can be submitted in one of the following categories:

Category-1: An embedded implementation on TI CC2652 MCU or a compatible IoT device

  • Submissions in this category will be evaluated based on classification accuracy, execution time per activity, memory requirements, and energy consumption per activity. The evaluation will be done with data that was not provided to the participants. This will ensure that all submissions are evaluated fairly.

Category-2: An FPGA implementation

  • Designs in this category will be evaluated based on classification accuracy, execution time, memory requirements, and area. The evaluation will be done with data that was not provided to the participants. This will ensure that all submissions are evaluated fairly.

All submissions must include

  1. One slide executive summary
  2. A two-page description of the design
  3. Their project source files (C, Verilog, or VHDL), including instructions to execute the code. Participants must also provide a test bench to execute the design

Presentation at the Conference (tentative)


The proposed contest will feature both presentations and demonstrations of the top designs in each category. The tentative outline is provided below:

  1. Introduction of the Contest (10 min)
  2. Background on Human Activity Recognition (10 min)
  3. Announcements of the winners in each category (5 min)
  4. Presentations of the selected entries (45 min)
  5. Demonstrations of selected designs (20 min)

Dates


Registration to the Contest:  May 25, 2022 – July 1st, 2022 (tentative)

Design Submission Deadline: August 31, 2022

Notifications: September 15, 2022

References


[1] J.-F. Daneault, “Could Wearable and Mobile Technology Improve the Management of Essential Tremor?” Frontiers in Neurology, vol. 9, pp. 257:1–257:8, 2018.

[2] A. J. Espay, et al., “Technology in Parkinson’s Disease: Challenges and Opportunities,” Movement Disorders, vol. 31, no. 9, pp. 1272–1282, 2016.

[3] G. Bhat, R. Deb, and U. Y. Ogras, “OpenHealth: Open Source Platform for Wearable Health Monitoring,” in IEEE Design & Test, March 2019, doi: 10.1109/MDAT.2019.2906110

[4] G. Bhat, R. Deb, V. V. Chaurasia, H. Shill, and U. Y. Ogras. “Online Human Activity Recognition using Low-Power Wearable Devices”, in Proc. of Intl. Conf. on Computer-Aided Design (ICCAD), Nov. 2018.

Chair


Ganapati Bhat
Embedded Human Activity Recognition Student Competition Chair
Washington State University
Website