SPRADE2 October   2023 AM69A

 

  1.   1
  2.   Abstract
  3. 1Introduction
  4. 2Localization and Mapping
    1. 2.1 Simultaneous Localization and Mapping
    2. 2.2 Graph SLAM
    3. 2.3 Localization
  5. 3Surroundings Perception
  6. 4Path Planning
  7. 5Summary

2.3 Localization

GUID-20230918-SS0I-LTX8-HVWP-PFC3NQ6TD5MM-low.svgFigure 2-3 Localization Process With Pre-built Map

Figure 2-3 shows the localization process in the mapped environment. This process is similar to the SLAM front end in Figure 2-2. The only difference is that, once the features are extracted from a frame, the corresponding features are searched from the map instead of ones from other frames. After finding the matched features, the pose of the mobile robot can be calculated by Perspective-n-Point (PnP), Iterative Closest Point (ICP), and so forth.

Table 2-1 summarizes the widely used technique in each step of the graph SLAM and localization.

Table 2-1 Various Techniques in Graph SLAM
Visual SLAMLiDAR SLAM
Feature extraction
  • Feature detection and descriptor:
    • SIFT, SURF, KAZE, AKAZE, ORB, BRISK
    • Convolutional Neural Network (CNN) based
  • 2D or 3D points itself
  • Geometric features
    • Edge
    • Plane
Feature association
  • Minimum Euclidean or Hamming distance between feature descriptors
  • Minimum Euclidean distance between 2D or 3D points, edges, and planes
Pose estimation
  • Direct Linear Transform (DLT)
  • LM optimization
  • PnP
  • Scan matching
    • Occupancy grid matching
    • ICP
    • Normal Distribution Transform (NDT)
  • LDT, LM optimization
Loop closure detection
  • Bag of Word (BoW)
  • Appearance-based global image descriptor
  • CNN based
  • Scan matching
  • Segmentation matching
  • CNN based
Graph optimization
(Bundle adjustment)
  • Make use of matrix sparsity
  • Newton, Gauss-Newton, LM underneath
  • Make use of matrix sparsity
  • Newton, Gauss-Newton, LM underneath

The AM69A embedded processor is an excellent choice for SLAM and localization. Octal A72 cores provide more than enough computing power for complex SLAM and localization algorithms. Many open-source algorithms can be quickly implemented and benchmarked on the AM69A. Moreover, the functional blocks such as feature extraction, feature matching, and pose estimation in Figure 2-2 and Figure 2-3 can be offloaded to hardware accelerators (HWAs) and C7x DSP to improve performance. Internal studies show that the throughput of the ORB SLAM with stereo camera is improved by 2 to 3 times by offloading stereo rectification and feature extraction to LDC, MSC, and DSP.