This is an old revision of the document!

Dataset of Surveillance Cameras : Tomo Kihara (Author) Github Repo for the model and the code —- What if we can train our own neural network to detect the presence of cameras that are watching us? With this question in mind, I started training neural networks that do image detection which would be able to detect the presence of a surveillance camera in the street. By using this function as a scavenger hunt-style game, the original idea was to create a game where players go in search of finding as many surveillance cameras on the street. Training and execution were done with tensorflow.js, which is a javascript version of tensorflow, a library dedicated to training neural networks. Tensorflow.js was used since it ran on smartphone web browsers which increase the accessibility of the game. Reinforcement learning was done with existing image detection model Mobilenet. During the process, 1045 images of a square and sphere-shaped surveillance cameras, fire extinguisher, the exit sign, and a bicycle were gathered and trained. 7 iteration was done to increase the accuracy of the categorization. In the end, the accuracy of more than 90% was achieved for each of the objects. However, when tested in the wild it had difficulty detecting the cameras due to the distance from the lens to the actual surveillance camera.

  • surveillance_camera.1565204635.txt.gz
  • Last modified: 2019/08/07 21:03
  • by waag