Cooking Activity Recognition using Deep Convolutional Bidirectional LSTM from Acceleration Data

Abstract

The difficulty of identifying a body's behavior based on sensor data, such as an accelerometer in a smartphone, is known as activity recognition. It's among the most widely studied topics in the field of machine learning-based classification. Cooking Activity Recognition Challenge (CARC) asked participants to recognize food preparation using motion capture and acceleration sensors. Two smartphones, two wristbands, and motion-capturing equipment were used to collect three-axis (x, y, z) acceleration data and motion data for the CARC dataset. One of the most challenging difficulties to solve in this investigation was identifying complicated tasks as smaller activities that are part of larger activities. Using a Convolutional Neural Network (CNN) and a Bidirectional LSTM, we’ve built a deep learning approach that extracts dynamical data for macro and micro activity identification. The model we proposed for that kind of dataset has a classification accuracy of 83% for macro activity and 85.3% for micro activity, respectively.