A narrow baseline increases the shared field of view of the two cameras, while yielding to shorter maximum range. Conversely, a larger baseline selleck decreases the common field of view, but leads to higher maximum range and accuracy at each visible distance. By employing the narrow baseline to reconstruct nearby points and the wide baseline for more distant points, the trinocular system takes the advantage of the small minimum range of the narrow baseline, while preserving, at the same time, the higher accuracy and maximum range of the wide baseline configuration. The trinocular system is integrated with a CLAAS AXION 840 4WD tractor (see Figure 1), which has been employed for the testing and the field validation of the system.
In Figure 1(b), the camera is visible, mounted on a frame attached to the vehicle’s body and tilted forward of about 12�� to minimize the field of view seeing the sky. The tractor’s sensor suite is completed by a 3D Sick laser rangefinder, a 94-GHz frequency modulated continuous wave (FMCW) radar, and a thermal infrared camera [7].Figure 1.The tractor test platform employed in this research (a), and its sensor suite (b).Table 1.Specifications of the stereovision system.The remainder of the paper is organized as follows. Section 2 reports related research in the field. The proposed self-learning framework is described in Section 3, whereas details of the statistical approach for ground classification are presented in Section 4. Sections 5 and 6 explain the geometry-based and color-based classifier, respectively.
In Section 7, the system is validated in field experiments performed with the tractor test platform. Section 8 concludes this paper.2.?Related WorkConsiderable progress has been made GSK-3 in recent years in designing autonomous, navigation systems for outdoor environments [8,9]. Progress has also been made in high-level scene analysis systems [10], with various application domains including on-road scene awareness [11,12], off-road rough terrain analysis for planetary rovers [13,14], off-road terrain classification for challenging vegetated areas [15,16], and agriculture [17�C19]. In this section, research is organized by its learning strategy: deterministic (no learning), supervised, and self-supervised. Estimating the traversability of the surrounding terrain constitutes an important part of the navigation problem, and deterministic solutions have been proposed by many.
However, deterministic selleck kinase inhibitor techniques assume that the characteristics of obstacles and traversable regions are fixed, and therefore they cannot easily adapt to changing environments [14,20,21]. Without learning, such systems are constrained to a limited range of predefined environments. A number of systems that incorporate supervised learning methods have also been proposed, many of them in the automotive field and for structured environments (road-following).