Abramson, Yotam (2005) AdaBoost/GA et filtrage particulaire: La vision par ordinateur au service de la sécurité routière. PhD thesis Informatique-Robotique, CAOR - Centre de robotique, ENSMP p.226.
Full text available as:
|
|
Abstract
This thesis presents two ITS applications,which are designed to be installed on a moving vehicle and detect other road users, using a single frontal camera. The two apllications are Stop and Go ACC and Pedestrian impact prediction.
The thesis opens by describing the history and current status of the ITS domain. We review several existing systems which represent several approaches and research directions. Among these systems there are ones which are operational or almost operational, and ones which are futuristic.
Next we present some novel results in the field of computer vision/machine learning. Thes results are using, and are partly motivated by, the example of pedestrian detection. In particular we present new type of weak-classifiers to be learned by the AdaBoost algorithm, a classifier which is working faster than others and is not dependant of scene lighting conditions. We also present a novel way to collect large high-quality training sets in order to vastly improve the training results.
Using thes results, we present a Stop and Go adaptive cruise control (ACC°. We implemented this application with a set of known image processing algorithms, demonstrating how the combination of several relatively-simple algorithms can yield a reliable system. The application is running in 10 images per second and follows the car in front, while using a motion estimator to detect cut-ins.
Our second application is a pedestrian detection and impact prediction application. The system is running in 10 image per second and reliably predict the probability of an impact with a pedestrian in some time frame.
| Item Type: | PhD Thesis (PhD) |
|---|---|
| Thesis Supervisor: | Laurgeau, Claude |
| Date: | 07 December 2005 |
| Board of examiners: | Laurgeau, Claude and Siarry, Patrick and Trassoudaine, Laurent and Steux, Bruno and Gallenne, Marie-Line and Milleman, Sylvain and Doyen, Philippe |
| Discipline: | Informatique-Robotique |
| Collection (Fonds): | ENSMP |
| Institution: | ENSMP |
| Department: | CAOR - Centre de robotique |
| Subjects: | 2. Information and Communication Sciences and Technologies |
| Uncontrolled Keywords: | Computer vision, Intelligent transportation systems, stop and go ACC, Pedestrian detection, Vision artificielle, Systèmes de transport intelligent, Détection de piétons, Commande de croisière adaptative. |
Table of content
Remerciements - i
Publications dans le cadre de cette thèse - iii
Résumé - iv
1 Introduction - 1
1.1 Etat de l'art - 1
1.2 Description des applications - 3
1.3 Contribution théorique - 4
1.4 Plan de la thèse - 4
1.5 Background - 4
1.6 Description of applications - 6
1.7 Theoretical contribution - 6
1.8 Program of this thesis - 7
2 Intelligent Transportation Systems (ITS) - 8
2.1 What is ITS? - 8
2.1.1 Introduction - 8
2.1.2 Historical background - 8
2.1.3 ITS Categories - 9
2.2 Infrastucture-based ITS - 10
2.2.1 Informative infrastucture systems - 10
2.2.2 The smart junction - 11
2.3 IV independent applications - 14
2.3.1 On-board GPS-based localization system - 14
2.3.2 On-board signalization - 15
2.3.3 Pedestrian detection - 16
2.3.4 Lane departure warning (LDW) - 17
2.3.5 Forward Collision Warning (FCW) - 19
2.3.6 Adaptive cruise control (ACC) - 20
2.3.7 Non-cooperative lateral control - 22
2.4 Vehicle-Infrastucture cooperative systems - 23
2.4.1 The AHS project - 24
2.4.2 The deployment problem - 26
2.5 Choise of applications - 28
2.5.1 List of possible applications - 28
2.5.2 Application quotation - 29
2.6 Choice of implementation methods - 31
2.6.1 Requirements for general detection and tracking applications - 31
2.6.2 Previous approaches - 31
2.6.3 Particle filtering - 32
2.6.4 The need for motion estimation - 32
2.6.5 The need for a learning algorithm - 33
2.7 Conclusion - 34
3 Visual object detection with machine learning - 35
3.1 Conventions and definitions - 35
3.1.1 Motion information - 35
3.1.2 Detection window - 35
3.1.3 Measuring the detection rate of a detector - 36
3.2 State of the art - 37
3.2.1 Support vector machine (SVM) - 37
3.2.2 Wavelets features for visual object detection - 40
3.2.3 AdaBoost - 44
3.2.4 Training a detector with AdaBoost - 47
3.3 The control-points features - 56
3.3.1 Independency of illumination - 58
3.3.2 Initial experiments - 60
3.3.3 Genetic algorithm as a week learner - 63
3.3.4 Feature discussion - 63
3.3.5 Using the attentional cascade - 67
3.4 The 5X5 moving kernel features - 68
3.4.1 Definition - 69
3.4.2 Initial experiments - 70
3.5 Analysis of features on synthetic data - 70
3.5.1 Noise sensitivity - 70
3.5.2 Sensitivity to hiding - 76
3.5.3 Conclusions - 77
3.6 Analysis of features on real data - 81
3.6.1 Without normalization - 81
3.6.2 With normalization - 82
3.7 SEVILLE: SEmi-automatic VIsuaL LEarning - 83
3.7.1 The setup - 83
3.7.2 The design of the experiment - 86
3.7.3 Theoretical justification - 87
3.7.4 The experiment - 89
3.7.5 Experimental results - 92
3.7.6 Observation at the separation process - 94
3.7.7 Conclusions and future work - 95
3.8 Comparison between various camera types - 96
3.8.1 Training details - 96
3.8.2 Regular camera at daylight time - 97
3.8.3 Near Infra Red (NIR) camera - 97
3.8.4 Far Infra Red (FIR) camera - 98
3.8.5 Conclusions - 98
3.9 Analysis of the cascade - 98
3.9.1 2D-ROC analysis - 99
3.9.2 3D-ROC analysis - 99
3.9.3 Conclusions - 100
3.10 Conclusions - 100
3.11 Future work - 100
3.11.1 The image pyramid- 101
3.11.2 Spatial density - 101
3.11.3 Tracking (temporal density) - 101
3.11.4 Low-level motion - 101
3.11.5 Limited search area - 101
3.11.6 Context learning - 102
4 Motion estimation algorithms - 111
4.1 Background - 111
4.2 general overview of the algorithm - 111
4.2.1 Motion models - 111
4.2.2 Other definitions - 112
4.2.3 Basic diagram - 112
4.3 Main (internal) loop on blocks - 112
4.3.1 Fetching of the new blocks - 112
4.3.2 Motion-model checking - 114
4.3.3 The actual SAD calculation - 114
4.3.4 The uniqueness of a motion model - 115
4.3.5 Minimal SAD values - 117
4.3.6 Spatial candidates generator - 117
4.3.7 Initiated candidates generator (uniqueness contradictor) - 117
4.3.8 Weights for different types of candidates - 118
4.4 External loop on images - 118
4.4.1 Grouping of areas - 118
4.4.2 High level information in grouping - 119
4.4.3 Minimal motion value - 119
4.4.4 Smoothing of objects - 119
4.4.5 Parameters calculation - 119
4.4.6 Temporal candidates generating - 120
4.4.7 Object based candidates - 120
4.4.8 Randomization - 120
4.5 Object tracking - 121
4.5.1 General method - 121
4.5.2 Upper and bottom confidence thresholds - 121
4.6 Algorithm future work-directions and risks - 122
4.6.1 Extended motion models - 122
4.7 Uniqueness of motion models - 122
4.7.1 Motion estimation for object detection and tracking - 124
4.7.2 Simple uniqueness - 124
4.7.3 Weighted uniqueness - 124
4.7.4 Towards one dimensional uniqueness - 125
4.7.5 Uniqueness as statistical dispersion - 125
4.7.6 one dimensional uniqueness as the weighted least squares - 126
4.7.7 Using the one dimensional uniqueness to enhance detection - 126
4.8 Conclusion - 126
5. Particles filter - 127
5.1 Introduction - 127
5.2 Nonlinear bayesian Tracking - 128
5.3 Optimal Algoritms - 129
5.3.1 Kalman Filter - 129
5.3.2 Grid-based Methods - 131
5.4 Particles Filter - 132
5.4.1 Sequential Importance Sampling (SIS) Algorithm - 132
5.4.2 Resampling and the generic particles filter - 133
5.4.3 Sequential Importance Resampling (SIR) Algorithm - 134
5.5 Conclusion - 136
6 Application I: Stop and Go system - 137
6.1 The detection and tracking framework - 137
6.1.1 Requirements for the application - 137
6.1.2 Previous approaches - 138
6.1.3 Particle filtering - 138
6.1.4 The state space - 141
6.1.5 How algorithms are combined - 142
6.1.6 Advantages of using particle filtering - 143
6.1.7 Confidence of targets - 143
6.2 Image processing algoritms - 144
6.2.1 Introduction - 144
6.2.2 Vehicle shadow detection - 146
6.2.3 Car's rear-lights detection - 152
6.2.4 Symmetry detection - 157
6.2.5 Vertical edges detection - 160
6.2.6 Detection with AdaBoost - 160
6.2.7 Weakness of algoriths - 167
6.3 Additional algoriths - 167
6.3.1 Motion segmentation as a tool for fast detection of cut-ins - 167
6.3.2 Integration with LIDAR - 168
6.4 The Stop and Go system - 177
6.4.1 System architecture - 177
6.4.2 Identifying the Stop and Go target - 178
6.4.3 Automatic disconnection of the system - 178
6.5 Results - 181
6.5.1 Typical examples of the particles filtering - 181
6.5.2 Analysis on a long sequence - 181
6.6 Conclusion - 188
7 Application II: pedestrian detection and impact prediction - 189
7.1 SEVILLE-based applications - 189
7.1.1 The SEVILLE pedestrian detector - 190
7.2 Particle-filter-based applications - 193
7.2.1 Introduction - 193
7.2.2 Pedestrien representation - 193
7.2.3 The basic algorithms and the likelihood function - 193
7.2.4 The input of the algorithms - 195
7.2.5 The projection and retro-projection - 195
7.2.6 The particle filtering framework - 195
7.3 The medium-level algorithms (MLA) - 195
7.3.1 The legs/diagonal detection algorithm - 196
7.3.2 Learning algorithm using AdaBoost - 201
7.3.3 Body detection - 201
7.3.4 Motion detection - 203
7.3.5 Vertical edges excluding detection - 204
7.4 Impact prediction - 205
7.4.1 Introduction - 205
7.4.2 Calculation of pedestrian's movement vector - 206
7.4.3 Impact prediction mechanism - 206
7.4.4 Using a pedestrian model - 207
7.5 Results - 211
7.5.1 Results on several examples - 211
7.5.2 Summarized results - 213
7.6 Conclusion - 213
8 Conclusion - 215
8.1 Vision artificielle - 215
8.2 Applications dans le domaine de l'automobile - 216
8.3 Perspectives - 217
8.4 Computer vision - 217
8.5 Automotive applications - 218
8.6 Future work - 218
Références bibliographiques - 220
Index - 225
| ID Code: | 1606 |
|---|---|
| Deposited By: | Céline Gueguen |
| Deposited On: | 10 March 2006 |
Repository Staff Only: edit this item