Stixmentation - Probabilistic Stixel based Traffic Scene Labeling

The detection of moving objects like vehicles, pedestrians or bicycles from a mobile platform is one of the most challenging and most important tasks for driver assistance and safety systems. For this purpose, we present a multi-class traffic scene segmentation approach based on the Dynamic Stixel World, an efficient super-pixel object representation. In this approach, each Stixel is assigned either to a quantized maneuver motion class like oncoming, or left-moving or to static background. The formulation integrates multiple 3D and motion features as well as spatio-temporal prior knowledge in a probabilistic conditional random field (CRF) framework. The real-time capable method is evaluated quantitatively in various challenging, cluttered urban traffic scenes. The experimental results yield highly accurate segmentation of urban traffic scenarios without the need for any manual parameter adjustments.