- Human Activity Recognition (HAR) is of great value in a wide range of real-world applications, such as health, fall detection and smart home.
- There are generally two types of solutions for HAR: device-based and device-free.
- Device-based solutions rely mostly on wearable devices such as smartphones and smart watches. However, these solutions often cause discomfort and extra burden.
- To overcome the weaknesses, device-free solutions utilizing cameras and Radio-Frequency (RF) signals have later come into view.
- Unlike camera based solutions, RF-based approaches do not raise privacy concerns, and are not affected by temperature or lighting conditions.
- Therefore, RF-based solution has become a promising candidate for indoor HAR, leading to a large amount of research contributions recently
- Original dataset was acquired from [1].
- Three different sensors and a total of 11 different activities and ambulatory gaits were considered, as listed in the Figure 1.
- The images from the respective sensors were resized to 128x128.
- Then the data from the respective sensors was splitted into train,validation and test sets.
- Model was trained on Google’s Collaboratory Platform for 20 epochs.
- Categorical Cross Entropy was loss function used along with Adam optimizer and Accuracy as metrics.
- Model gave accuracy of 99% on training dataset and 89% on validation dataset.
- The Classification Model was further optimized for edge inference using Float-16 Quantization Technique.
- The weights of the classification model were converted 16-bit floating point values as part of post training Quantization process using TensorFlow Lite.
- The 6×reduction in model size from 15.1 MB to 2.5 MB was observed post quantization.
- It also made the model computationally faster for execution.
