4.2 Obstacle classification using laser information
4.2.2 SVM classification
SVM classification is performed using the SVM implementation from the Computer Vision OpenCV library. SVM algorithm was developed by Vapnik and Cortes in [89] and is widely used in machine learning as a classification
method. SVM computes features from the positive and negative samples used for training. In images, HOG features and LBP features are common, but laser point clouds are not suitable per se for SVM classification. A number of mathematical features have been defined in order to extract from the point clouds the required information about the shape ans characteristics of the detected obstacle, and to keep the information in a fixed size regardless the size of the point cloud, as required for SVM training and classification.
Laser scanner clusters feature vector
Clusters detected in laser scanner generated point clouds are used to determine a Region of Interest in the image where we can perform obstacle classification applying Computer Vision and Artificial Intelligence techniques, but can also be used for obstacle classification without image support [12]. Clusters are converted into a mesh structure by Delaunay triangulation in order to recon-struct the shape of the obstacle and to extract relevant features according to the 3d shape of the cluster, as seen in figure 42. The mesh can be represented from any point of view; figure 42a represents the view from the laser, and 42b represents a aerial view of a point cloud.
(a) Frontal view (b) Aerial view Fig. 42 Mesh representation of a cluster.
These obstacles are detected by the system as clusters, which have some characteristics suitable for further SVM training following the process out-lined in figure 43. Clusters obtained from the test sequences are stored and manually labeled using the corresponding images for training. These clusters
4.2 Obstacle classification using laser information 79
are manually labeled as frontal view, back view, side view, frontal oblique view and back oblique view.
Fig. 43 SVM learning process for clusters: Training and classification.
Previous works as [14] have considered 2D point clouds for classification, but the present work is intended to extract features from a 3D point cloud, in an effort to maximize the use of the available information. Some of the features considered are described in Table 4.
Table 4 Some of the features considered for cluster classification.
Feature Meaning
Concentration Normalized mean distance to the centroid 3D
Y-Z concentration Normalized mean distance to the centroid excluding x X-Z concentration Normalized mean distance to the centroid excluding y X-Y concentration Normalized mean distance to the centroid excluding z Flatness Normalized mean distance to the most populated plane
found in the cluster
Sphericity Normalized mean distance to the most populated sphere
Cubicity Measures how far are the planes containing the mesh triangles from being the same plane or from being perpendicular
Triangularity Measures the uniformity of the triangles composing the mesh bye the relation between sides’ lengths Average deviation Average deviation from the median in x, y, z
A study of the relevance of every feature considered has been performed, using a set of training of 14,000 clusters representing a pedestrian and 8,400 clusters representing several kind of non-pedestrian obstacles. It is important to use only features that help to differentiate between positive and negative existence of the Object Of Interest. A similar study has been performed for car, bicycles and motorcycles clusters database used for classifier training, in order to select the appropriated features for each kind of obstacle.
In figure 44, several good features are studied. The horizontal axis indicates the number of sample considered, and the vertical axis represents the magni-tude of the feature. Red crosses represent the value of the positive samples, while green crosses are the values of the negative samples. The features describing well the difference between positive and negative samples present a high concentration of positive magnitudes, very different from the negative magnitudes, as shown in figure 44e, cluster width, with the positive values very concentrated near zero, and negative magnitudes concentrated higher
4.2 Obstacle classification using laser information 81
than zero. The rest of the examples in figure 44 show highly differentiated values for positive and negative samples.
0 5000 10000 15000
(b) Mean distance to closest point in cluster
(c) Mean distance to farthest point in cluster
(d) Cluster height * depth * width
(f) Cluster width * height
Fig. 44 Distribution of the values of a feature describing well a cluster characteristic.
Figure 45 represent statistics for features describing poorly the difference between positive and negative samples. Figure 45a shows that the feature Cluster density from upper viewdoes not define well the difference between positive and negative samples, as most of the positive and negative values for that particular feature are coincident.
0 5000 10000 15000
(a) Cluster density from upper view
(b) Number of points in cluster
0 5000 10000 15000
Fig. 45 Distribution of the values of a feature describing poorly a cluster characteristic.
Chapter 5
Obstacle detection and classification using computer vision
After the initial laser-based obstacle detection and classification stage, obsta-cles represented as clusters in the laser scanner Point Cloud are translated into Regions Of Interest (ROI) in the image, where Objects Of Interest (OOI) are searched using Computer Vision algorithms, as will be shown in the present chapter. Although public datasets are available for pedestrians and cars, like INRIA, ETH, TUD-Brussels, Daimler, Daimler stereo, Caltech-USA and KITTI, experience shows that best results for classification are achieved when the same camera and in the same position is used for training and for classifi-cation. Having this goal in mind, several datasets have been created for LSI using the XB3 camera.
5.1 LSI Datasets
Using computer vision for obstacle classification requires a dataset of positive and negative samples of the Objects Of Interest (OOI). Several public dataset are available, but better results are expected if a dataset obtained from the same sensors, located in the same position of the vehicle is used. For this reason,
image datasets for pedestrian and bicycles have been created at Intelligent Systems Lab (LSI) as part of the present thesis.