husainattar / rail-rush

Keras based implementation of the CSRNET model

Rail-Rush

It is the Python implement Computer vision model which is used in the Rail-Rush App to detect crowd density in public transportation.

For the application code base check out this repo

This is a Keras with tensorflow backend Implementation.We have Implemented a CSRNET model with a slight modification by using slight help from MCNN model. We have trained our Model on Shanghai Dataset with the Links provided below.

CSRNET Original Paper

MCNN Paper

Shanghai-Dataset

Understanding the Dataset

We have used the .mat file provided and converted into density map. Preprocesing process in the paper and in our implementation is almost same with slight modification. For better understanding of the dataset and mat file check out the link provided.

Mat File Understanding.

We have used K-nearest node algorithm for finding the closest nodes rather than going with linear search.

We have also used the multiplying factor of 0.1 rather than provided in the paper.

Models , Working and Testing

As stated above we have follow the paper but with slight modification by adding the batch_Normalize layer for efficient result.

Before using this technique we have also used HOG_MODEL and HAAR_CASCADE_MODEL and their implementation is also provided in the repository.

We have generated 2 models from dataset using the CSRNET technique one for sparse and other for dense crowd.

We have tested this model on Videos on platform and generated the result.