Vision-based parking assistance system for leaving perpendicular andangle parking lots

2013/12/17

指導教授 : 張元翔 老師研究生 : 林柏維通訊碩一 10279026

1

Outline

• Introduction

• Related work

• Methods A.STHOL feature selection

B.Bayesian decision

• Results

• Conclusion and Future

2

Outline

• <Introduction>

• Related work

• Methods A.STHOL feature selection

B.Bayesian decision

• Results

• Conclusion and Future

3

Introduction(1)

• In the last years, a considerable number of research works and industrial developments on Intelligent Parking Assist Systems (IPAS) have been proposed, including both assistance and automatic parking approaches.

4

Introduction(2)

• The main goal of IPAS that assist drivers when parking is to ease the maneuver avoiding small collisions, reducing car damage, and avoiding personal injuries.

• In this paper a new vision-based Advanced Driver Assistance System (ADAS) is proposed to deal with scenarios like the ones depicted in Figs. 1(a) and 1(b).

5

Introduction(3)

6

(a) Back-out perpendicular parking

(b) Back-out angle parking

Fig. 1. Driver and camera Field of View (FOV).

Outline

• Introduction

• <Related work>

• Methods A.STHOL feature selection

B.Bayesian decision

• Results

• Conclusion and Future

7

Related work(1)

• The proposed architecture of the system is composed of three main parts: camera, processor and CAN-Bus communications.

• From the CAN-Bus we obtain the next variables: steering angle, car speed and current gear.

8

Related work(2)

• These variables are used to trigger on/off the detection module according to the Finite State Machine (FSM).

• Then the system waits until the car has been put into reverse gear and the detection module is then triggered on.

9

Related work(3)

10

Fig. 2. FSM for detection module.

Related work(4)

11

Fig. 3. Camera located at the back-right side of the vehicle.

Related work(5)

12

(a) Driver’s point of view

(b) Image captured by the camera

Fig. 4. Driver and camera point of view.

Outline

• Introduction

• Related work

• <Methods > A.STHOL feature selection

B.Bayesian decision

• Results

• Conclusion and Future

13

Methods

• The spatio-temporal domain is analyzed by accounting the number of lines and their length with respect to their orientations in a histogram of orientations that we so-called Spatio-Temporal Histograms of Oriented Lines (STHOL).

14

Flow chart

15

Fig. 5. Overview of the spatio-temporal detection module.

Spatio-temporal images

16

Fig. 6. Two examples of the scan-lines and spatio temporal images.

Outline

• Introduction

• Related work

• Methods <A.STHOL feature selection>

B.Bayesian decision

• Results

• Conclusion and Future

17

STHOL feature selection(1)

• The first step of the line detection is the computation of the image derivatives using Sobel edge detector.

• The edge pixels having the same label are then grouped together using connected components algorithm.

18

STHOL feature selection(2)

• The line segment candidates are obtained by fitting a line parameterized by an angle q and a distance from the origin p using the following expression:

p= x cos(q) +ysin(q) (1)

19

STHOL feature selection(3)

• The line parameters are then determined from the eigenvalues and and eigenvectors v1 and v2 of the matrix D associated with the line support region which is given by:

20

STHOL feature selection(4)

• If the eigenvector v1 is associated with the largest eigenvalue, the line parameters (p,q ) are determined using:

21

STHOL feature selection(5)

22

Fig. 7. Overview of the STHOL feature selection architecture.

Outline

• Introduction

• Related work

• Methods A.STHOL feature selection

<B.Bayesian decision>

• Results

• Conclusion and Future

23

Bayesian decision(1)

• we represent the image I in terms of STHOL features and follow a Bayesian approach considering the free traffic class:

24

Bayesian decision(2)

• Use the minimum-error-rate classification using the discriminant function:

25

Bayesian decision(3)

• Instead of using two discriminant functions and , and assigning to we define a single discriminant function and we finally trigger the warning signal if .

26

Outline

• Introduction

• Related work

• Methods A.STHOL feature selection

B.Bayesian decision

• <Results>

• Conclusion and Future

27

Results

28

Fig. 8. Three sample sequences with the triggered warning signal.

Outline

• Introduction

• Related work

• Methods A.STHOL feature selection

B.Bayesian decision

• Results

• <Conclusion and Future>

29

Conclusion and Future

• This paper presented a novel solution to a new type of ADAS to automatically warn the driver when backing out in perpendicular or angle parking lots.

• A novel spatio-temporal motion descriptor is presented(STHOL features) to robustly represent oncoming traffic or free traffic states.

• A Bayesian framework is finally used to trigger the warning signal.

30

Conclusion and Future(2)

• Future work will be concerned with the evaluation of the method in night-time conditions, comparisons between our generative approach and other discriminative approaches such as SVM-based.

31

Thank you for your listening

32