Occupancy grids have been widely used in various areas including robotics. For example, suppose you are developing an autonomous vehicle, one of the first thing you will probability do is to buy a LiDar sensor and attach it to your car. From the sensory inputs, you can generate a two-dimensional occupancy map, which is often referred to as an occupancy grid map.
1. We have an occupancy grid map. (binary map)
Unlike structured environments such as industrial factories, things are quiet hard to predict what's going on on the roads. Some drivers might be drunk and act abnormally or some pedestrians may jaywalk right in front of you. Perhaps, due to the bad weather conditions, you may not be able to see clearly, or far enough. In such cases, relying on visual information may not be a good idea.
2. Predicting future is very very important.
3. Handling noise is crucial!
(Can you differentiate from obstacles to background noises?)
4. Visual information may not be reliable.
5. How can we infer flow directions as well as how future occupancy grids will change.
(Flow directions are shown with different colors)
Summarizing what I've been talking about, I am currently developing an algorithm that can predict future occupancy information using a sequence of occupancy maps. I call my algorithm as occupancy flow method as it infers the directional flow information using a variant of recurrent neural network architecture. It has a two-layer architecture where the first layer works as a temporal denoising autoencoder, and the second layer predicts future occupancies as well as flow directions.
The synthetic simulation results are shown under.
The top-right input (100 by 100 binary matrix) is given as an input to the recurrent network.
The outputs of first and second layer are shown in bottom-left and right, respectively.
Despite extreme salt and pepper noises, the final output in the second layer successfully capture the occupancy flow as well as predict how the occupancy grids will change.
Video
'Enginius > Machine Learning' 카테고리의 다른 글
| 강화 학습이란? (2) | 2017.05.24 |
|---|---|
| Total variational distance between two empirical distributions (KDE) (0) | 2016.06.06 |
| Bayesian Optimization with Expected Improvement (0) | 2015.12.16 |
| Convolutional Neural Network (CNN) (5) | 2015.12.03 |
| Deep Neural Network (DNN) (1) | 2015.12.02 |
