This project aims to detect objects in a video stream then generate data that can be consumed by various clients e.g. SPA apps, chat bots etc.

It comprises of a python script that uses pretrained tensorflow models for out of the box inference.

The system is conceived as micro services which will communicate via publish/subscribe queues.

A source could be a video file, a network stream, a directory of images. The service that reads the source is also aware of the instance name from the configuration. The source creates a message with a unique ID.

The source pushes the message into a channel where services that communicate with model services (models) receive messages.

At the time of pushing the message, a shared counter, say counter -in, with identified by the message ID is created with the total number of model services that received the message for processing.

After each model service produces some detection results, those results are placed on a publish/subscribe channel where all related results are aggregated. The aggregation can be handled by via another shared counter, say counter -out which is incremented by each model service upon receiving the message from the source.

Once the two counters are equal, the message is placed on another broadcast channel where other services take the message and deliver it appropriately.

A conda environment is created first and when activated, additional pip packages are installed.

• Run conda env create -f env.yaml to setup a conda environment
• activate environment source activate object_detection_mini
• install pip packages pip install -r requirements.txt --no-deps

• setup and start the tensor flow serving docker container

  First time run:

  sudo docker run --name ssd_mobilenet_v1_coco -p 8501:8501 -p 8500:8500 --mount type=bind,source=/home/mugo/downloaded-tensorflow-models/ssd_mobilenet_v1_coco_2017_11_17/,target=/models/ssd_mobilenet_v1_coco -e MODEL_NAME=ssd_mobilenet_v1_coco -t tensorflow/serving

  Every other run:

  sudo docker start -a ssd_mobilenet_v1_coco

  Stopping:

  sudo docker stop ssd_mobilenet_v1_coco

• Run video detection script

bash run_with_env.sh python detect_video_stream_tf_serving.py ~/Videos/train-passenger-foot-stuck.mp4 ~/tensorflow-models-repo/research/object_detection/data/mscoco_complete_label_map.pbtxt 8500 ssd_mobilenet_v1_coco predictions --cutoff 70

Individual tests can be run like this:

bash run_with_env.sh pytest detect_video_stream_utils_test.py --disable-warnings --log-cli-level=DEBUG

• https://github.com/kunadawa/object-detection-event-web-server
• https://github.com/kunadawa/object-detection-react-app
• https://github.com/ericnjogu/object-detection-protos

• https://towardsdatascience.com/detecting-pikachu-in-videos-using-tensorflow-object-detection-cd872ac42c1d
• https://github.com/tensorflow/models/blob/master/research/object_detection/object_detection_tutorial.ipynb
• https://www.semantics3.com/blog/a-simplified-guide-to-grpc-in-python-6c4e25f0c506/
• https://medium.com/innovation-machine/deploying-object-detection-model-with-tensorflow-serving-part-3-6a3d59c1e7c0
• https://medium.com/@yuu.ishikawa/how-to-show-signatures-of-tensorflow-saved-model-5ac56cf1960f
• other credits are given in code and commit comments.

MIT

Eric Njogu