Moving Object Detection

 

 

 

Research  

 

Object detection and tracking

Moving Object Detection

Background modeling techniques for foreground moving object detection.

- High-quality nonparametric moving object detection implemented in real-time in a GPU.

- Spatio-temporal background and foreground models.

- Selective update of the background model.

- Foreground modeling augmented with a tracking system based on a particle filter capable of dealing with a variable and unknown number of moving regions.

- Bayesian classifier that is able to combine models with different spatial widths and to add pixel-level prior data for each model.

- Selective analysis to automatically select regions of interest in the input images.

- Robustness against shadows cast by moving objects.

  

(a) Example of fission of groups of particles.

(b) Example of fusion of groups of particles.

 

Example of prior foreground probability distribution

 

Selective analysis of a frame. Subfigure (a) shows an image featuring an already established object and a new object just entering the frame on the left that has not been deteted. Subfigure (b) shows, iteration by iteration, the detection process where the new object is correctly detected; blue blocks are those to be analyzed in the next iteration and red/green blocks are those that tested negative/positive. We can see in the zeroth iteration that the blocks where the already established object is expected to be located are marked for analysis from the beginning and in the first iteration the new object is hit by the random blue blocks, correctly growing to full detection in the next iterations. Subfigure (c) shows the current image and the final detection. 

Results

 

Results on SABS database. Color notation: correct detections in green, false detections in black and misdetections in red.

 

Results on LASIESTA database. Color notation: correct detections in green, false detections in black and misdetections in red.

 

Results in SABS database. Color notation: correct detections in green, false detections in black and misdetections in red.