Independent Study: Sanders
Sensor fusion methods for object detection/ scene understanding in Outdoor Mobile Autonomous Robots
Research Objective
-
To identify key features that required sensor input for navigation
-
To find out sensor types/ fusion methods need for KanGo (An outdoor autonomous ground robot) to achieve basic scene understanding or object detection.
-
To discover and compare sensors and methods
Context/ Assumption
The scope of the research would be focusing on the Perception part of the navigation, which the required input and algorithm is too complex and large in scale in this study.
Insights
-
Camera is the one of the most suitable type of sensors for object classification as similar human, optics information comes with high accurate. Existing algorithms for edge detections and neural networks like CNN have already been mature enough to be launched in the industry.
-
There are few types of camera we can still select from:
-
Typical Cameras like, Monocular Camera, Stereo Cameras, RGB-D Cameras or Omni-directional Cameras
-
the quality of the image highly depends on the light setting of the environment which makes them perform worse at night.
-
-
LiDAR Cameras
-
In that way, compared to normal cameras, the accuracy is higher and it is almost unaffected by the global light settings. The cost of the LiDar sensor is cost :), given that it is more expensive and has more power consumption.
-
-
Event Cameras
-
be able to obtain the edge of a moving object at night. There are many researchers able to recreate an actual image of the scene, as the camera moves with the robot itself. Secondly, the power consumption of event cameras is generally lower than the typical ones. Although it is not cheap currently, the are researchers creating DVS that cost less e.g. using FPGAs.
-
Proposed Fusion Method to estimate object size, location and speed.
Significance of Research
While Navigation being a great part in the system of the project,
The study provide the fundamental scenery understanding for the robot to decide on its next move according to the different situations.
A simple example would be whether the robot waits for the obstacle to pass, or the robot would need to change its local path to pass the obstacle.