Object tracking for sediment transport analysis

Collaboration with ETNA group of Irstea Grenoble

In this work, we focus on bedload sediment transport experiments in a turbulent flow were sediments are represented by small spherical calibrated glass beads. The aim is to track all beads over very long time sequences (~ 1 million images) to obtain sediment velocities and concentration.

The main bottlenecks to solve are:

1. Diversity and abruptness of motion – Three different motion types are possible: resting when a bead is not moving, rolling when it is rolling near the bed surface and saltating when it is bouncing on others above the bed surface. A given bead can suddenly switch from one motion type to another depending on its velocity, neighborhood and interactions with other beads.

2. Weakness of the transparent beads detector – Due to 3D effects, the bead pattern is slightly changing resulting in miss-detections and lack of precision in the detection of transparent beads. Combined with the abruptness of motion, this constraint can lead to wrong associations and bad trajectories

Our contribution is to propose a multiple motion particle filter (MMPF) based on mechanical dynamics and a non-linear observation model to handle miss-detections. It uses the technique of switching dynamical model and a graded observation through a detector confidence estimation.

The code implementing the tracking algorithms[1],[2] is available here.

Evaluations were made using two different datasets with their associated ground-truth (the datasets can be downloaded here ) :

- Experimental sequence – It was obtained using the experimental flume of Irstea Grenoble to study the collective behavior of beads with respect to their size. It consists of à 1,000-frame sequence recorded at 130 fps

- Numerical sequence – The idea is to reproduce as best as possible our experiments by numerical simulation. This 10,000 frame sequence was generated thanks to a model developed at Irstea based on a coupled fluid discrete element method.

Compared to the state-of-the-art, our algorithm provides the highest tracking precision and accuracy. Moreover, it appears to be the least impacted by missing detections.


[1] H. Lafaye de Micheaux, C. Ducottet, P. Frey, Online multi-model particle filter-based tracking to study bedload transport, ICIP 2016, IEEE International Conference on Image Processing, pp. 3489-3493, 2016.

[2] H. Lafaye de Micheaux, C. Ducottet, P. Frey, Multi-model particle filter-based tracking with switching dynamical state to study bedload transport, Machine Vision and Applications, Vol. 29(5), pp. 735-747, 2018

< Back