uo-object_counting API

based on python-flask-docker-sklearn-template
https://github.com/Soluto/python-flask-sklearn-docker-template
Simple websockets + flask + tensorflow + sqlite API providing counts of road users in still images available via UTMC public API https://www.netraveldata.co.uk/?page_id=9\ Organised and sorted images can be accessed at Urban Observatory UTMC CCTV portal https://api.newcastle.urbanobservatory.ac.uk/camera/

Websockets is listening for data from 'UTMC Open Camera Feeds' on wss://api.newcastle.urbanobservatory.ac.uk/stream
Image URL's are passed to (included) object detection API (i.p.add.ress:port/detection/api/v1.0/count_objects?img_url=some_address_to_image)
The object detection API returns counts of objects of different types (defined by tensorflow model and associated labels)
Camera name, timestamp (both extracted from images URL), image URL and object detection API's response are stored in sqlite file database Application's parametrization is done via environmental variables when container is started (or can be included in Dockerfile)
Two included Dockerfiles (with_gpu and without_gpu) can be used depending on available resources

SETUP

Build

build for use with or without GPU
docker build . -f Dockerfile-with_gpu -t tflsk_gpu
docker build . -f Dockerfile-without_gpu -t tflsk_nogpu

Starting

without GPU
docker run -d --rm -v /folder/to/store/sqlite/filedatabase:/some/place e DB='/some/place/cctv_counts.db' -e ENVIRONMENT='production' -e MODEL='fig_frcnn_rebuscov-3.pb' -e LABELS='rebuscov-classes-3.pbtxt' -e GPU_MEMORY=1 -e MIN_CONF=0.33 -e W=640 -e H=480 -p 6001:80 tflsk_nogpu

running with GPU requires nvidia-docker rather than regular docker (setting runtime depends on nvidia-docker version, either "--runtime=nvidia" or "--gpus all")

docker run -d --gpus all --rm -e DB_NAME=uo-vision-1 DB_USER=uo DB_PASS=vision DB_DOMAIN=database.urbanobservatory.ac.uk DB_PORT=5324 -e ENVIRONMENT='production' -e IP="172.17.0.2" -e MODEL='fig_frcnn_rebuscov-3.pb' -e LABELS='rebuscov-classes-3.pbtxt' -e GPU_MEMORY=0.25 -e MIN_CONF=0.33 -e W=640 -e H=480 -p 6001:80 --name STILLS_COUNTER tflsk_gpu

docker run -d --restart=always --gpus all -e DB_NAME=uo-vision-1 -e DB_USER=uo -e DB_PASS=vision -e DB_DOMAIN=database.urbanobservatory.ac.uk -e DB_PORT=5234 -e ENVIRONMENT='production' -e MODEL='fig_frcnn_rebuscov-3.pb' -e LABELS='rebuscov-classes-3.pbtxt' -e GPU_MEMORY=0.25 -e MIN_CONF=0.25 -e W=640 -e H=480 -e CUDA_VISIBLE_DEVICES=0 -p 6001:80 tflsk_gpu

Parameters and environmental variables:\

Use a mounted volume -v to assure data persistance
DB path to file database, use in tandem with mounted volume
ENVIRONMENT production/local switches port between 80/5000
MODEL path to detection model (relative to main.py)
LABELS path to object classes mapping
GPU_MEMORY what fraction of memory to use for this application (set 1 for non-gpu machine to not try and override defaults which in some cases may cause errors)
MIN_CONF minimum confidence score for object detector to accept object proposal and count it
W and H width and height for resizing images before feeding them into detection model

docker run -d --restart=always --gpus device=0 -v /home/ntk38/:/Storage -e STORAGE="/Storage" -e DB_NAME=uo-vision-1 -e DB_USER=uo -e DB_PASS=vision -e DB_DOMAIN=database.urbanobservatory.ac.uk -e DB_PORT=5234 -e ENVIRONMENT='production' -e MODEL='fig_frcnn_rebuscov-3.pb' -e LABELS='rebuscov-classes-3.pbtxt' -e GPU_MEMORY=0.25 -e MIN_CONF=0.25 -e W=640 -e H=480 -e CUDA_VISIBLE_DEVICES=0 -p 6001:80 --name STILLS_COUNTER tflsk_gpu

Usage

Object detection endpoint

call API to get counts of different types of objects in an image provide via 'image_url' parameter
127.0.0.1:6001/detection/api/v1.0/count_objects?img_url=https://file.newcastle.urbanobservatory.ac.uk/camera-feeds/GH_A692B1/20200319/191336.jpg
respose will be e.g.
{"car": {"count": 11}, "person": {"count": 1}, "van": {"count": 1}, "truck": {"count": 1}}

another call can request counts of objects in specified camera camera= or all cameras when supplied parameter is 0 (zero). Other parameter is minutes than will tell the API how far back from now you want to get the data for.

127.0.0.1:6001/db/api/v1.0/get_counts?camera=0&minutes=60
response will be e.g.
[{"counts": "{"car": {"count": 3}}", "camera": "GH_A167G1:V02", "url": "https://file.newcastle.urbanobservatory.ac.uk/camera-feeds/GH_A167G1/20200324/175515.jpg", "ts": "2020-03-24 17:55:15"}, {"counts": "{"car": {"count": 9}}", "camera": "GH_A167I1:V01", "url": "https://file.newcastle.urbanobservatory.ac.uk/camera-feeds/GH_A167I1/20200324/175516.jpg", "ts": "2020-03-24 17:55:16"}, {"counts": "{"car": {"count": 4}}", "camera": "GH_A167K1:V01", "url": "https://file.newcastle.urbanobservatory.ac.uk/camera-feeds/GH_A167K1/20200324/175526.jpg", "ts": "2020-03-24 17:55:26"}, {"counts": "{"car": {"count": 3}}", "camera": "GH_A184A2:V01", "url": "https://file.newcastle.urbanobservatory.ac.uk/camera-feeds/GH_A184A2/20200324/175557.jpg", "ts": "2020-03-24 17:55:57"}

... or if specifying camera (e.g. NC_B1318B1):

127.0.0.1:6001/db/api/v1.0/get_counts?camera=NC_B1318B1&minutes=5 response:
[{"counts": "{}", "camera": "NC_B1318B1:V01", "url": "https://file.newcastle.urbanobservatory.ac.uk/camera-feeds/NC_B1318B1/20200324/182942.jpg", "ts": "2020-03-24 18:29:42"}, {"counts": "{"car": {"count": 1}}", "camera": "NC_B1318B1:V02", "url": "https://file.newcastle.urbanobservatory.ac.uk/camera-feeds/NC_B1318B1/20200324/183149.jpg", "ts": "2020-03-24 18:31:49"}]

Note that in response camera name has view name appended after :, this information comes from api.newcastle.urbanobservatory.ac.uk/stream provided by a separate app (image clustering) https://github.com/urbanobservatory/uo-cam2views

Object Detection Model

The model included in /app/ folder is FRCNN 640x480 trained on PitchIn Rebuscov dataset (collection of ~10K traffic CCTV images), publicly available soon...

TomKomar / uo-object_counting