Counting-people-video

From the CCTV footage in each room, we get the number of people currently standing in there and generate insights about the inflow of visitors throughout the day. This way you could concentrate on the areas where visitors are not going.

Using the Tensorflow Object detection API, we will be counting the number of people in a video. A frame is extracted every 30 seconds from the video and a forward pass of the model is performed. If a person is found in the video, then the count is increased.

Requirements

OpenCV - 3.3.1
Tornado
Tensorflow
Protocol buffer compiler

Installation instructions

# For CPU
pip install tensorflow
# For GPU
pip install tensorflow-gpu

For Ubuntu

sudo apt-get install protobuf-compiler 

For OSX

brew install protobuf

Other Libraries

pip install opencv-python
pip install tornado # For running the server 

Tensorflow object detection API

protoc utils/*.proto --python_out=.

Running

If you want to test out the implementaion then you can use the object_detect.py

python object_detect.py --path <path to video>

To run the server

python server.py

Instruction to plot bounding boxes

As per the origial implementation of the tensorflow object detection API, the bounding boxes are normalised. To get the original dimensions you need to do the following.

(left, right, top, bottom) = (xmin * im_width, xmax * im_width,
                              ymin * im_height, ymax * im_height)