Real-time occupancy grid mapping using LSD-SLAM

[摘要] ENGLISH ABSTRACT: This thesis investigates the use of semi-dense depth data from monocular vision using large scale direct SLAM (LSD-SLAM) to create accurate occupancy grid maps for autonomous navigation, in real-time. Having an accurate map is crucial for an autonomous system to avoid collisions and remain safewithin its environment. Sensors used to gather information on the environmentare typically associated with some degree of uncertainty, and this must be considered when building a map. An autonomously navigating system also needs to have clear definition of free and occupied space within its environment.Literature shows that LSD-SLAM has great potential as a highly accurate and real-time SLAM algorithm; however, the resulting map is in the form of a semi-dense point-cloud which is not immediately useful to an autonomouslynavigating system. The point-cloud map must therefore be processed further.Occupancy grid maps (OGMs) offer an ideal solution for map representation that is useful for autonomous navigation. The environment is divided intoevenly spaced grid cells, each representing a probability of occupancy. OGMs also allow maps to be efficiently updated with the incorporation of uncertaintyfrom sensor measurements.Inverse sensor models (ISMs) can be used to characterise the uncertainty of a particular sensor and to calculate the prediction of occupancy given asensor measurement and its uncertainty. A literature review shows that two popular ISMs (one by Thrun and one by Andert) can be used in conjunctionwith the depth estimates of LSD-SLAM to create an OGM. Literature also shows that each of these ISMs contains a parameter that is associated with very little information on how their values should be chosen, and we thereforeincluded this in our investigation.We design a mapping system using the aforementioned ISMs, which runs in parallel with LSD-SLAM. Initial tests show that the performance of theopen-source version of LSD-SLAM did not agree with the author's claims.The results also revealed a significant lack of sufficient datasets for our main evaluations on map accuracy.Our mapping system was tested on 3 main criteria: memory usage, performanceand accuracy. All evaluations were performed on both ISMs, on various datasets, over a range resolutions and parameter changes for each ISM. Resultsshowed that Thrun's ISM out-performed Andert's ISM on all criteria, and that our system could indeed produce accurate maps that could be useful for autonomousnavigation. The results also showed that the 'default choice for theparameters of each ISM is not necessarily always sufficient. Additionally, we conclude that LSD-SLAM does not perform well in terms of 30 Hz real-timerequirements, while our mapping system can.

[发布机构] Stellenbosch University