Using the Analytic Feature Framework for the Detection of Occluded Objects

In this paper we apply the analytic feature framework, which was originally proposed for the large scale identification of segmented objects, for object detection in complex traffic scenes. We describe the necessary adaptations and show the competitiveness of the framework on different real-world data sets. Similar to the current state-of-the-art, the evaluation reveals a strong degradation of performance with increasing occlusion of the objects. We shortly discuss possible steps to tackle this problem and numerically analyze typical occlusion cases for a car detection task. Motivated by the fact that most cars are occluded by other cars, we present first promising results for a framework that uses separate classifiers for unoccluded and occluded cars and takes their mutual response characteristic into account. This training procedure can be applied to many other trainable detection approaches.