2022 IB Diploma Extended Essays

Figure 3. The figure above demonstrates sensor fusion architecture. By combining information from the LiDAR, Camera and the IMU. Therefore correcting the orientation and position of the robot from all different sensors. Then combining that information together through EKF. It creates a much more accurate position and orientation compared too Hector SLAM, ORB2-SLAM and ZED fusion [6]. This occurs firstly because by combining information from the IMU, it makes the slam algorithms less susceptible too fast changes in acceleration, velocity and rotation, which the IMU can correct [6][7]. Secondly, combining information from the Hector Mapping with ORB2- SLAM data can give a more accurate position and orientation of the robot [8]. This is because Hector SLAM provides depth of every feature within a scene, however ORB-SLAM uses a sparse method of extracting data, where it picks specific features within a scene in order to localise itself.[9] Furthermore, ORB-SLAM incorporates the bag-of-words (BOW) model. This method can create a feature map of the environment in real-time stably in many scenarios. Loop detection and closing via BOW is effectively prevented as the cumulative error can be quickly retrieved after the tracking is lost. [9] Similarly, Hector Mapping uses the Gauss-Newton method to solve the problem of scan matching, this method does not need odemetry method and such can be quite reliable [10]. Lastly, by combining data from ZEDfu the robot can get a complete 3D scene map and therefore localise itself within 3D dimension rather than 2D [11]. This is because ZED-fu utilises the RGB-D camera which can provide both colour and depth information in its view field. It is the most capable sensor for building a complete 3D scene map [11]. This method (zed-fu) combines the intensity error and depth error of pixels as error functions, and minimises these cost function to obtain the optimal camera pose. This process is implemented through g2o. Entropy-based key frame extraction and closed-loop detection method are further utilised, thus greatly reducing the path error of the robot [12].