Developing intelligent crowd analysis pipeline for hospitals (JIPMER IITM collaboration project).
The paper we are primarily implementing is resnet crowd.
Using tensorflow-keras for implementation.
Created TfRecord files for the input data and parsed using Tensorflow Data-API.
Used ShanghaiTech dataset for the training purpose.
Did multi-task learning by truncating layers from resnet50 by adding two branches of network.
Performed transfer learning using imagenet weights, freezing the resnet50 part of the model and training only the added branches.
Using AdaGrad Optimizer and L2 loss for both
heatmap and count estimation.
- Heatmaps from images.
- Count from images.
- Violent activity recognition.
- Crowd movement prediction.
- Locating and tracking of abnormal detected region.
- Person re-identification