CODEWITHSUNDEEP
tensorflowcomputer-vision
Eugenio
Eugenio

Reputation: 681

Home surveillance, TensorFlow (+ OpenCV)

I am considering to implement a network of security cameras in my countryside property due to the recent intrusion of some burglars. To do so, I am trying to make the most out of the equipment I already have at home, such as:

I wanted to feed the video stream of each camera "into" TensorFlow+OpenCV (or other DeepLearning algorithms) to make sure I can recognise what's/who's in my garden, garage, etc. - basically it would be:

planning to use Telegram or other APIs provided by the instant messaging platform to be notified real-time.

Configuring Tensorflow on my ubuntu 16.04 LTS (workstation with GPU support) was not trivial at all nor it will be the process of recognising objects and people.

Do you have any recommendation on this little project? Would it make sense using Docker to make the configuration smoother?

Upvotes: 2

Views: 758

Answers (1)

Nouman Riaz Khan
Nouman Riaz Khan

Reputation: 259

You can break it down as follow:

1 - Decide between bounding box and semantic segmentation

2 - Select your pre-trained model after selecting your framework (Tensorflow in your case). You need to beat around Real Time vs Accuracy 2.1 - The best in my opinion would be MobilenetV2 based deeplab (segmentation) or NASNET (bounding box)

3 - Retrain it for transfer learning if Cars & Persons are not in your model, otherwise you dont need to

4 - Look for model optimization if required for Jetson TK1 (compression etc)

5 - Once successfully deployed and it starts looking for person and cars, develop your logic for how should it alert you. (e.g. absence of car for x consecutive frames or presence of person for x consecutive frames etc)

Happy Surveillaying!

Upvotes: 2

Related Questions