Human Activity Recognition using Convolution Neural Networks

The given repository builds an efficient code for action recognition (humans) using UCF101 with the help of Keras. Convolution Neural Networks are used for training the dataset. Also, accuracies are compared for different pre-trained models trained on Imagenet dataset. Videos can be viewed as a sequences of images(frames), thats why they can be processed in ways similar to images.

Dataset

UCF101 is an action recognition data set of realistic action videos, collected from YouTube, having 101 action categories. It consists a total of 13,320 videos from 101 actions.

The video preprocessing has been done using OpenCV. Each videos is 25 fps. To reduce the amount of redundant data to be fed into the network, 1 frame per second from each video is extracted which later on is fed into the network model. FFmpeg can also be used for preprocessing the video data into frames.

Pre-trained Models

The Following models were used for training the dataset. Also, image size was altered to check the dependencies.

• VGG16
• InceptionV3
• DenseNet

Training and Testing

Training

Make sure to have the requirements installed as provided in the requirements.txt.

pip install -r requirements.txt

Execute DV/train.py file in your preferred python execution setup.

The best model outputs will be automatically saved with the code provided in the file itself,.

Testing

The code provided in the DV/model_pred.py file is used for testing.

Test Outputs

The summary of different models trained are given below along with different parameters.

Pre-trained model	Batch Size	Image Dimensions	Testing Accuracy score
VGG16	128	128 * 128 * 3	45.20
VGG16	128	200 * 200 * 3	41.50
VGG16	128	32 * 32 * 3	27.62
VGG16	128	64 * 64 * 3	37.69
VGG16	128	150 * 150 * 3	48.40
VGG16	256	150 * 150 * 3	50.14
VGG16	128	150 * 150 * 3	42.40
VGG16	128	150 * 150 * 3	48.26

Pre-trained model	Batch Size	Image Dimensions	Testing Accuracy score
Inception V3	256	150 * 150 * 3	59.08
Inception V3	256	200 * 200 * 3	61.88
Inception V3	256	200 * 200 * 3	53.23
DenseNet 121	256	200 * 200 * 3	70.39
DenseNet 169	256	200 * 200 * 3	70.39
DenseNet 201	256	200 * 200 * 3	73.27

Deployment

Django

The model had been deployed using Django web Framework. To run the model on localhost run:

python manage.py runserver

Some screenshots for the running model are given below. (Firefox and Chrome are recommended for running the project smoothly.)

Acknowledgement

• A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014.
• X. Peng, L. Wang, X. Wang, and Y. Qiao. Bag of visual words and fusion methods for action recognition: Comprehensive study and good practice. CoRR, 2014.