Light-Field Camera Based Spatial Multiple Small Particle Tracking with Post-processing

Multi-object detection and tracking is one of the most significant tasks in computer vision, which has been widely applied in various fields, such as biomedicine, automatic driving, and video surveillance. In our project, we aim to investigate the flight behavior of different small refuse-derived fuel particles (e.g., plastic, wood, paper, etc.) with sizes around 5 to 30 mm, which are blown into a combustion chamber (rotary kiln) via a lance, by detecting and tracking the particles. By utilizing a calibrated light-field camera, both 2D gray value images and 3D information, including the depth of the scene, are available for investigation. To obtain particle trajectories, accurate temporal localization of the particles is necessary. Thus, we perform an approach combining a 2D detection method and a 3D clustering method for precise detection of the particles at first. Subsequently, the temporal 2D particle localizations are associated into tracklets based on the Kalman filter state estimator and the global nearest neighbor algorithm. Owing to the wrong and incomplete tracklets caused by several inaccurate detections, we present a post-processing method to enhance the tracking performance in the paper. ... mehrFinally, the optimized 2D particle trajectories are converted into 3D trajectories, and these obtained 3D particle trajectories are analyzed by a polynomial regression for further analysis of the particle flight behavior.