cirquit
PhD student with a focus on machine learning, distributed systems and functional programming.
SLAM With Kalman Filtered Odometry in O(1)
Abschlussvortrag zur Masterarbeit
Aufgabensteller: Prof. Dr. Kranzlmüller
Betreuer: Tobias Fuchs
Datum des Vortrags: 21.03.2019
Autonomous Driving
1
Real track
Seen track
UAS Munich Team
2
Combustion
Since 2006
Electrical
Since 2010
Driverless
Since 2017
Interdisciplinary team of 100+ students from UAS Munich & LMU
Simulatenous Localization and Mapping
3
Sensors
Localization
Mapping
Estimated Position
Environmental Map
Autonomous Pipeline
4
Visual Sensor
Neural Network
SLAM
Trajectory
Planning
Velocity Sensor
Accelerometer
Vehicle
Control Unit
Sensors
Preprocessing
Planning
Excecution
Master Thesis
Problem Statement
5
How can heterogeneous sensor inputs be combined in a SLAM process component with constant overhead and improved precision?
SLAM Design - Localization
Yaw Rate (rad/s) EKF (4x1)
- Kistler
- Visual Odometry
Velocity (m/s) EKF (4x2)
- BOSCH
- BOSCH + Kistler
- POSIROT + Kistler
- Single-track vehicle model
6
Visual Odometry
7
SLAM Design - Localization - Vel. EKF
8
SLAM Design - Localization - Yaw Rate EKF
9
SLAM Design - Mapping
10
SLAM Design - Mapping
- Maximum distance between cones
- We need to see two cone pairs to recognize curves
- The vehicle can only move forward
- The length of the track is capped at 500m
- We know the framerate of the object detection
- We control how fast the vehicle goes
Possible preallocation
$$\mathcal{O}(N^2) \rightarrow\mathcal{O}(N)$$
$$\mathcal{O}(N) \rightarrow\mathcal{O}(1)$$
11
SLAM Design - Mapping
12
Evaluation - Positional Error
13
Evaluation - Yaw Rate EKF
14
Difference to EKF Yaw Rate
Evaluation - Velocity EKF
15
Evaluation - Space Complexity
16
valgrind
summary of the offline logging
Evaluation - Time Complexity
17
$$\mu = 609 \text{ Hz}, \;\;\;\; \sigma^2 = 82 \text{ Hz}$$
Evaluation - Time Complexity
18
$$\mathcal{O}(1)$$
$$\mathcal{O}(N)$$
logging
without
logging
$$\mu = 30433 \text{ Hz}, \;\;\;\; \sigma^2 = 6305 \text{ Hz}$$
$$\mu = 32488 \text{ Hz}, \;\;\;\; \sigma^2 = 8032 \text{ Hz}$$
$$\mu = 193061 \text{ Hz}$$
$$\sigma^2 = 36483 \text{ Hz}$$
$$\mu = 183201 \text{ Hz}$$
$$\sigma^2 = 54198 \text{ Hz}$$
Summary
19
- Stable clock frequency
- Enable preemptive scheudling in the Linux kernel
- UDP Logging in a custom thread
- Developed with OpenMP in mind
- Enhanced estimation
- Particle Filter
- More precise object detector
Future Work
Findings
- Sensor Fusion model with
- EKF
- Gaussian Mixture Models
- Showed feasibility of SLAM
- Proof of concept
$$\mathcal{O}(1)$$
Backup #1: Related Work
20
MonoSLAM
ORB-SLAM2
FastSLAM
- only indoor use
- maximum amount of features before slowdown
$$\mathcal{O}(N^2)$$
- no approximation for computational complexity
- computational complexity is fixed at
$$\mathcal{O}(N \log N)$$
Backup #2: Related Work
21
Backup #3: Data Assoc. Calculations
22
Backup #4: Hardware Overview
23
hard realtime
soft realtime
Backup #5: Cars Overview
24
FSAustria
2017
Business Plan
FSAustria
2017
Acceleration
FSGermany
2017
Gesamt
#3
#1
#1
FSItaly
2018
FSGermany
2018
Acceleration
FSGermany
2018
Gesamt
#5
#2
#2
Gesamt
Acceleration 0-100km/h: 2.1s
Logging from 02.08.2018
- 635 seconds (~ 10 min)
- 2.5 m/s
- 7 laps
Extracted Values
- Objectcount
- distance (m), angle (rad), color (yellow, blue, red)
- Velocities (vx, vy) in m/s from Kistler sensor
- Velocities (vx, vy) from visual odometry
- Acceleration (ax, ay) in m/s^2 from BOSCH sensor
- Yaw rate (rad/s) from BOSCH sensor
- Yaw rate (rad/s) from Single-track vehicle model
- Yaw rate (rad/s) from POSIROT + Kistler
- Yaw rate (rad/s) from BOSCH + Kistler
- Steering angle (rad) from POSIROT sensor
- Delta time (s) since last tick
Backup #6: Experimental Setup
25
SLAM with Kalman filtered odometry in O(1)
By cirquit
SLAM with Kalman filtered odometry in O(1)
Abschlussvortrag
