Visual Appearance Analysis of Forest Scenes for Monocular SLAM

Abstract

Monocular simultaneous localisation and mapping (SLAM) is a cheap and energy efficient way to enable Unmanned Aerial Vehicles (UAVs) to safely navigate managed forests and gather data crucial for monitoring tree health. SLAM research, however, has mostly been conducted in structured human environments, and as such is poorly adapted to unstructured forests. In this paper, we compare the performance of state of the art monocular SLAM systems on forest data and use visual appearance statistics to characterise the differences between forests and other environments, including a photorealistic simulated forest. We find that SLAM systems struggle with all but the most straightforward forest terrain and identify key attributes (lighting changes and in-scene motion) which distinguish forest scenes from "classic" urban datasets. These differences offer an insight into what makes forests harder to map and open the way for targeted improvements. We also demonstrate that even simulations that look impressive to the human eye can fail to properly reflect the difficult attributes of the environment they simulate, and provide suggestions for more closely mimicking natural scenes.

Lost in the Woods? Place Recognition for Navigation in Difficult Forest Environments

“…The track at the start of the SFU videos must be difficult to distinguish, likely due to most frames consisting solely of a road and surrounding treeline. To the eye these images are very similar, as noted in Garforth and Webb (2019). Under the canopy, however, the more complex skyline has been shown to be useful in navigation before (Stone et al, 2016), so we posit that this is part of what the network is using to achieve its improved performance.…”

“…The simulated forest data is classified significantly less accurately than real data, with the most common non-forest labels being pond, fishpond and aquarium (which account for 23.63% of classifications between them). The artificial lighting conditions seem the most likely explanation for this, which is not unsurprising given the results of Garforth and Webb (2019) and furthers their warning about using simulated data to test algorithms meant to work in the real world.…”

“…The likely cause of these difficulties is the very highly textured scenes providing an abundance of potential features, or the notable change in lighting that occurs even in the few minutes between the images in Figure 7, such that repeated visits to the same location do not have a large enough overlap in extracted features to find an alignment. This was a risk, given the previously reported poor performance of SLAM systems in these environments in Garforth and Webb (2019), and shows that more work must be done on the frame to frame tracking part of the mapping process before a full working forest SLAM system can be produced.…”

“…Natural environments like forests present a series of problems for traditional visual SLAM systems, as outlined in Garforth and Webb ( 2019 ). Firstly, they undergo frequent rapid changes in illumination due to the effect of the canopy blocking parts of the bright sky.…”

“…Mobile robotic systems have the potential to aid in forest management by improving the efficiency, scale and accuracy of tree health data gathering. In order to do this, robots must be able to navigate autonomously around large GPS-denied forest environments where modern visual Simultaneous Localization and Mapping (SLAM) techniques often struggle (Garforth and Webb, 2019 ).…”