Overview / Usage

This project demonstrates the use of deep learning to recognize human activity from a live/stored video stream.

Methodology / Approach

The project uses a filtered version (top-20 classes) of the **UCF101 **dataset. The dataset is prepared in the following manner:

• Individual frames are extracted from the videos and are serialized as well as the information about the labels (activity label) of these individual frames.

After these frames and their labels are prepared, they are fed to a deep learning model for training and during prediction time, the idea of rolling averaging is used.

For deep learning, I used transfer learning i.e. I fine-tuned the VGG16 network (the top portion of it) and trained it on the filtered dataset. Now, the aim is to further optimize the model using OpenVINO and run that on a combination of RPi + NCS2.

Technologies Used

• Keras
• OpenCV
• Intel Python
• GCP