Section: New Results

Feature Selection for road obstacles classification

Participants : Itheri Yahiaoui, Pierre Merdrignac, Anne Verroust-Blondet.

In order to ensure the ability of an automated vehicle to be autonomous in a real environment, we must equip it with tools (hardware and software) to meet the requirement of such an application as safety, real-time processing, understanding and intelligence, etc. To contribute to these objectives a perception system is of vital importance. The one on road obstacles detection and classification is of particular interest for us. In this work a large number of geometric features have been proposed to describe different class objects like vehicles, pedestrians, cyclists and static obstacles from 2D laser points. A binary classification was performed with an Adaboost algorithm. In order to improve this work and enhance the classification rate, we have constructed new binary and multiclass classifiers, using SVM and logistic regression, with optimal choices of kernel parameters and models. We have defined several decision strategies by tracking objects in the video sequences, which lead to obtain the most probable target object. On the other hand, we have studied different dependence measures between the proposed features and the classes, leading the selection of the best set of features. As measures of dependence, we have used nonparametric estimate of mutual information, Fisher information and Pearson correlation. We have used also the Akaïke criterion in order to select the best models (the best subset of features) in logistic regression.