Stixel World Based Long-Term Object Tracking for Intelligent Driving

Long-term object tracking is key for a higher level of semantic interpretation of driving environment. One of the state-of-the-art approaches for long-term tracking is Tracking-Learning-Detection (TLD), which, however, suffers from variability of on-road objects and moving cluttered background. This paper presents a long-term object tracking method for intelligent driving based on Stixel World to address the drifting problem. First, this method adopts TLD framework, and integrates intensity and depth cues into the detector and learning component. Next, this method introduces Stixel World for compact medium-level representation of the 3D world, and Attention Guiding Filter (AGF) is proposed to focus on relevant areas in the image. Experiments in real traffic scene show the outstanding long-term tracking performance for intelligent driving.