An ultra-fast human detection method for color-depth camera

• An extremely fast technique to locate positions that are plausibly humans is proposed to quickly reduce searching space.
• A novel knowledge based human locator which can deal with partial occlusion and incomplete depth data is proposed.
• A stepwise filtering framework enables the system to perform very quickly (80–140 fps).

Real-time human detection is important for a wide range of applications. The task is highly challenging due to occlusions, complex backgrounds, and variation of human poses. We propose a cascade-structured approach to real-time human detection in cluttered and dynamic environments with both color and depth data seamlessly incorporated. The first stage efficiently exploits depth data which generates a set of physically plausible yet over-detected candidates. These candidates are then purified by another two filters: a knowledge based human upper portion locator and a data-driven learning based filter. Experimental results show high detection accuracy achieved by the proposed method at 80–140 fps on a single CPU core (without GPU acceleration).