Skip to content

Latest commit

 

History

History
72 lines (57 loc) · 2.91 KB

Work2_log.md

File metadata and controls

72 lines (57 loc) · 2.91 KB
Raw

Part2 Reachability Analysis

Input Uncertainty

  • Method1: Analytical

  • Method2: Gazebo experiemental

  • Method3: real-experiemental (also as validition for other methods)

  • Method_New1: sensor measurement error + detection algorithm err from gazebo (simulate perfect sensor in gazebo)

  • To merge 2 gaussian distribution

  • Method1 (not feasible)

    • Sensor measurement error:

      • Kinect Depth resolution: ~ 1.5 mm at 50 cm. About 5 cm at 5 m Noise: About +-1 DN at all depths. This means +-1 mm close, and +- 5 cm far.

      • Ouster OS-1 16-Zeilen RANGE: 120 m PRECISION: ±1.5 – 10 cm

    • Error from detectors (e.g. leg_detecot/ upper_body_detecor)

      • Purpose1: detector algorithm itself does not generate localization error? since it's only doing recognition
        • to verify: In gazebo with single detector and perfect sensor

    • Error from sensor-fusion

      • Sensor fusion matches detection from 2 detectors
      • when matched take mean-pose as new fused_pose
      • Covariance is set to mean of 2 covariances too. (Not how it supposed to be)

  • Method2, details of experiment setup In gazebo sensor are simulated as perfect (No noise)

    1. Single detector (leg/upper) a. static single person to get static localisation err [3*3 locations] b. Single moving person [Normal] c. Single moving person [Slow] d. Single moving person [fast] e. multiple people Visulise results if its plausible [Idealwise pose should be perfect]
    2. Tracker [fused] b. Single moving person [Normal] c. Single moving person [Slow] d. Single moving person [fast] e. multiple people

  • Detector's uncertainty

    • Leg_detecot (DavidLu)
      • with real-time covariance, pos.covariance[0] = pow(0.3 / reliability, 2.0); ideal case cov = 0.09
      • If this is a experience-formel, has to be validated through tests

  • To Systematically analysis and summerize the uncertainty from gazebo-experient

    • A python script is written to record the data

      1. For single person, easy
      2. For multiple person, a match method is needed -> Benchmark provide help

    • Valid sensor range is essential for us

      • In frame: robot base
        • RGB-D:
          • depth 2.2 ~ 6.4 # Change plugin parameter of point cloud cut off 0.5 ~5
          • horizotal range +- 0.3 at 2.2, +-2,7 at 6.4
        • Laser:
          • depth 2 ~ 6,7
          • horizotal: +- 0.3 at 2, +-4 at 6, +-2.4 at 6.7
        • Laser high recall:
          • depth 2 ~ 15
          • horizotal: +-2 at 2, +-8 at 10 [almost the whole laser range]

    • Some comment to the citaion in seminar-report

    [45]: Meanly focused on the robustness of the Collision-waring system against the sensor err