5

I'm quite new to image processing and AI. But I have the expertise to create a network that can be used in object detection and recognition. Most of the time I've used ANN or Naive Bayes.

Now, I want to develop a method of action recognition, something like identifying whether one is jogging, running or walking by applying ANN. However, I really don't have idea how the sequence of frames can be classified.

In static image, segmentation and feature extraction is easy. But in regard to a moving image, I'm unsure of the approach.

Thanks in advance!

DukeZhou
  • 6,209
  • 5
  • 27
  • 54
Ramah Wart
  • 63
  • 3

2 Answers2

0

LSTM's provide a simple way to process sequential data (assuming all video sequences have the same number of frames and resolution).

DukeZhou
  • 6,209
  • 5
  • 27
  • 54
dant
  • 101
0

Depending on the complexity of the problem you could go with the full blown approach of LSTMs (like dant suggested) or you could try to address the challenge using pre-processing and stick with your already established CNN architecture.

If you want to use pre-processing, you basically calculate delta images between frames and use them like "normal" frames in your CNN. The DQN of Deepmind used this approach successfully to understand motion when playing ATARI games. Those images were rather simple and the amount of motion limited especially compared with videos of real life situations, so this simple approach may not be sufficient for your problem, but I consider it worth trying.

Demento
  • 1,684
  • 1
  • 9
  • 26