In my thesis, I developed an approach for estimating velocity vectors of both static and moving objects in RGB-D images, which is critical for applications such as obstacle avoidance where understanding the dynamics of a scene is crucial. This was accomplished by first preprocessing the RGB-D data to isolate the objects by removing the floor from the scenes, which simplifies the tracking and categorization of objects in the image.
I employed a Voxel Region Growing algorithm for clustering scene elements based on their spatial locations, which aids in differentiating and tracking objects across consecutive frames. To associate objects across these frames, I adapted the Hungarian algorithm, enabling the precise computation of their velocity vectors. This method ensures that each object's trajectory is maintained, which is essential for the subsequent velocity calculations.
The velocity vectors are computed and validated against a ground truth to verify their accuracy. This validation is essential for ensuring that the vectors can reliably be used in practical scenarios, such as navigating through dynamic environments where accurate real-time responses are necessary.
I implemented the entire algorithm in C++ to leverage its performance capabilities, achieving cycle times of approximately 8 ms at a resolution of 848x480 pixels. This efficiency allows for real-time processing, which is crucial for the deployment in robotic systems and other applications requiring immediate reaction to changing conditions.
My thesis presents a comprehensive method for real-time velocity estimation using RGB-D images. The approach focuses on efficient data processing techniques, accurate object tracking across frames, and robust velocity computation, all of which contribute to the reliability and practicality of obstacle avoidance systems.