Terrain Recognition and Classification based using Convolution Neural Networks

Project Title

Terrain Recognition and Classification based using Convolution Neural Networks

Project Area of Specialization Artificial IntelligenceProject Summary

Mobile robots frequently operate in rough, uneven terrains. In certain situations, the workspace is always rough terrain instead of flat ground. To maximize the safety degree of a path, robots should simultaneously find a relative flat path to the goal position and try to keep a safe distance from the danger sources. One way for them to identify easier to traverse paths is to use deep learning methods, such as a convolutional neural network (CNN).

The idea is to implement a convolution neural network-based terrain recognition system that allows different variety of robots (biped robots, legged robots, etc.) to find a safe path for their objective. The safe path might include lack of danger, travelable slope and various other factors that help the robot walk comfortably.

Project Objectives

The aim of this project is to:

• Retrieve the data from the environment using ToF Camera (discussed later) or from the terrain model with the help of RGB-D sensor (such as Microsoft Kinect, ASUS Xtron, etc.)
• To provide that data to the central processor (Raspberry Pi 3B+) to analyze the nature of the terrain.
• To extract features for Neural Network algorithm
• To develop an Artificial Neural Network algorithm
• Allow the robot to select the safest path

Project Implementation Method

For feature recognition of terrain, we’ll use a ToF (Time of Flight) camera (such as Microsoft Kinect v2, Intel RealSense D345, Argos 3D P100, etc.). A ToF camera uses infrared light (lasers invisible to human eyes) to determine depth information - a bit like how a bat senses its surroundings. The sensor emits a light signal, which hits the subject and returns to the sensor. The time it takes to bounce back is then measured and provides depth-mapping capabilities. This provides a huge advantage over other technologies, as it can accurately measure distances in a complete scene with a single laser pulse. The accuracy and precision of ToF Camera for their usage in the context of 3D reconstruction and SLAM (Simultaneous Localization and Mapping) allows the controller to adopt the data (distance, depth, etc.) from the environment to be gathered. We can input three channel RGB images, disparity maps, grayscale images, etc., or some combination of these formats to a convolution neural network such that it can make accurate predictions about the current terrain.

Benefits of the Project

Following are the benefits of the project:

•  Allows for Terrain Recognition and Classification
• Allows to predict the shortest route possible
• Can be implemented on to the industrial purposes to minimize manpower
• Development of Convolution Neural Network (CNN)
• Its accuracy is not limited to any certain point as it increases with each sample
• It is cost effective and practical in real time hardware implementation

Technical Details of Final Deliverable

Versatile robots as often as possible work in unpleasant, lopsided landscapes. In specific circumstances, the workspace is in every case harsh territory rather than level ground. To augment the wellbeing level of a way, robots should all the while locate a relative level way to the objective position and attempt to keep a sheltered good ways from the peril sources. One route for them to recognize simpler to navigate ways is to utilize profound learning techniques, for example, a convolutional neural system (CNN). One route for portable robots to distinguish simpler to cross ways is to utilize profound learning strategies, for example, a convolutional neural system (CNN). It isn't clear, in any case, what info ought to be given to the CNN to best empower it to order extraordinary territory. 

A landscape model can likewise be utilized to investigate the territory and to permit the robot to gain proficiency with the most secure and fastest way so as to finish their assignment. The sketch recovered from the 3D territory model can give all the important data required to comprehend the landscape.

Final Deliverable of the Project Hardware SystemCore Industry ITOther Industries IT Core Technology NeuroTechOther Technologies RoboticsSustainable Development Goals Industry, Innovation and InfrastructureRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	71000
Raspberry Pi 3B+	Equipment	1	7000	7000
Pi Camera v2	Miscellaneous	1	4000	4000
Microsoft Kinect v2	Equipment	1	60000	60000