Terrain classification using ToF sensors for the enhancement of agricultural machinery traversability

Thumbnail Image
Yandun, Francisco J.
Gregorio López, EduardGregorio López, Eduard - ORCID ID
Escolà i Agustí, AlexandreEscolà i Agustí, Alexandre - ORCID ID
Rosell Polo, Joan RamonRosell Polo, Joan Ramon - ORCID ID
Torres-Torriti, Miguel
Auat Cheein, Fernando A.
Other authors
cc-by-nc-nd, (c) Elsevier, 2017
Journal Title
Journal ISSN
Volume Title
Ground properties influence various aspects of mobile machinery navigation including localization, mobility status or task execution. Excessive slipping, skidding or trapping situations can compromise the vehicle itself or other elements in the workspace. Thus, detecting the soil surface characteristics is an important issue for performing different activities in an efficient, safe and satisfactory manner. In agricultural applications, this point is specially important since activities such as seeding, fertilizing, or ploughing are carried on within off-road landscapes which contain a diversity of terrains that modify the navigation behaviour of the vehicle. Thus, the machinery requires a cognitive capability to understand the surrounding terrain type or its characteristics in order to take the proper guidance or control actions. This work is focused on the soil surface classification by implementing a visual system capable to distinguish between five usual types of off-road terrains. Computer vision and machine learning techniques are applied to characterize the texture and color of images acquired with a Microsoft Kinect V2 sensor. In a first stage, development tests showed that only infra-red and RGB streams are useful to obtain satisfactory accuracy rates (above 90%). The second stage included field trials with the sensor mounted on a mobile robot driving through various agricultural landscapes. These scenarios did not present illumination restrictions nor ideal driving roads; hence, conditions can resemble real agricultural operations. In such circumstances, the proposed approach showed robustness and reliability, obtaining an average of 85.20% of successful classifications when tested along 17 trials within agricultural landscapes.
Journal or Serie
Journal of Terramechanics, 2017, vol. 76 (2018), p. 1-13