This repository consists of python scripts used for the task of detecting people within an office. This is done by capturing an overhead image, splitting the image into 5 smaller images and performing basic pre-processing on these 5 images. Once processed, these images are run through a trained CNN that classifies each image as containing people (or not). The result of this classification can be saved for assessing the CNN's accuracy or broadcast over the network as a binary vector (1's for occupied, else 0) for other applications.

Since the image of the room is split into 5 regions (4 corners and a middle region), the vector has a length of 5, with a 1 corresponding to an occupied region and 0 being unoccupied. The spatial positions corresponding to each vector index are as follows: [top left corner, top right corner, bottom left corner, bottom right corner, middle of room].

Within this repository, create a data/ folder that contains an image_mean.npy file, and the compiled computational graph of your trained Neural Network, named graph. The mean file can be generated using the convert_mean.py file, as explained below. The details of generating a graph file from a trained caffe model are outlined here.

convert_mean.py:
Converts Caffe's binaryproto files used for the mean image into the npy format, which can be used by the Movidius API.
Usage: Modify the hardcoded path for the mean binaryproto file within the program, and run the script (no args) to generate a local image_mean.npy file.

image_process.py:
Consists of image processing functions similar to those used in training of the Caffe Classification network. Note that the network requires a 227x227 image due to the constraints of transfer learning, and so the transform_img function is used for resizing the images and providing the necessary padding to the image.
Usage: Should be used as a library, see network_accuracy.py and occupancy_server.py for example usages of the functions.

network_accuracy.py:
Used to assess the accuracy of the CNN's predictions. Saves a captured image and its predicted occupancy label, which can then be validated manually to see the true accuracy of the network.
Usage: There are 4 main paths that should be supplied to the program. The HOME_PATH in the script is the location of this repository. The GRAPH_PATH should point to the data/graph file within the repository. The mean_img is the path to the image mean in the npy format. The final path that must be supplied is the output_path, which will be populated with labelled folders for each day containing an image and its corresponding vector label in a text file. The output_path is supplied as a command line argument when the script is run.
Note: Two variables may be useful for customizing the behaviour of the script.

• captureRate determines the period at which images are taken, in seconds.
• endTime is the time at which the script will end on the current day, which is useful if the script is launched automatically on a daily basis (with a cronjob)

occupancy_server.py:
Sends the predicted occupancy of the room as a binary vector.
Usage: As with network_accuracy.py, necessary paths should be modified in the script. By default, the occupancy server broadcasts an occupancy vector over port 1234 upon receiving a "Read" message (case-sensitive) from the client. Thus the client should connect to the Raspberry Pi's IP address and broadcast a "Read" message when a prediction is desired. No command-line arguments are required when running the script.