Experiments and results

In this page I propose different scenarios, the setup of various experiments, and I will show the results in terms of accuracy or demonstrative videos.

Scenarios

  • Scenario A consists of random events generated by the machine on a set of photos regarding a navigation session.
  • Scenario B involves the recognition of human legs among objects with similar shapes.
  • Scenario C is a multi-class classification, realized with three binary SVM models.
  • Scenario D is a small object recognition experiment to show the power and the limits of the detection system.
  • Scenario E regards a door handle detection

Results obtained in Gazebo simulations:

Scenario A experiment

Two videos as a demonstration of multi-class recognition task, i.e. Scenario C:

  1. The Turtlebot, in order to escape from a blind alley, decides to move the balls instead of the cube thanks to its recognition capabilities;
  2. In the second, in its path, the robot meets a series of obstacles, a wall and then a human being, and reacts every time in the most appropriate way possible.
wall-ball_cut.mp4
human-columns-wall_cut.mp4

Finally, in Scenario D, our purpose is the small object detection. The cascade architecture is a great help when it comes to identify small objects among all the pixels of an input image: observe how the accuracy increases from a one to a four layers cascade.

One layer cascade confusion matrix

Four layers cascade confusion matrix

Results obtained in a real environment:

The stress test that I prepared to assign a score is based in scenario B:

  • five different persons
  • two different heights for each person (1 meter and 50 centimeters)
  • four poses for each height (1: first pose near, 5: first pose far, 2: second pose near, 6: second pose far, and so on...


To evaluate the performance of the cascade in the case of scenario E, I prepared a stress test consisting on 60 images with the following characteristics:

  • five doors
  • 9 shots for each door (between brackets there is the legend for reading Sheet "ScenarioE")
  • three frontal with varying distances (1 for short, 2 for medium, 3 for far)
  • two not centered with medium distance (4 for right, 5 for left)
  • two not centered from below with short distance (6 for right, 7 for left)
  • two not centered from above with short distance (8 for right, 9 for left)
  • different light conditions
  • 15 images of walls that do not contain any handle

Scenario B: pose 1-5

pose 2-6

pose 3-7

pose 4-8

Real world results

The objective of Scenario D is to detect a small ball:

Integration with Learning by Demonstration Module

After the separated development of the situation detection model in this thesis and the learning-by-demonstration framework in an external thesis, it is time to join the two parts together. Thanks to the overall system, a robot is able to apprehend when and how take a decision that was not predicted in the original plan. This decision, which depends on the information acquired through the camera, enables an opportunistic behavior that may consist in a maneuver, a talking option or other case-specific actions.

The architecture, after the training phase, of the overall system used in the following experiments works in this way: the classifiers recognize different situations, then they communicate their decision to the PNP (Petri Net Plan) framework through a specific ROS topic. Depending on the message received, PNP enables one of the policy generator and the robot undertake the correspondent action.

Pepper_plant_fast.mp4
marrtino_human_box_fast.mp4