A self-organizing approach to detection of moving patterns for real-time applications

Lucia Maddalena, ICAR-CNR, Naples Branch
Alfredo Petrosino, University Parthenope of Naples

This page has been created in order to show results of moving object detection using algorithm HSV-SO presented in [1] on different image sequences, and to compare them with those obtained with other existing algorithms.

1. Sequence MSA

Size: 320*240
Source: home-made

We report two consecutive sequence frames (Previous and Actual frames) and the actual frame with corresponding moving object detection mask computed with HSV-SO algorithm. The detection mask shows that the walking man is perfectly detected. Moreover the bag, which has been left by the man in previous frames, is still detected as an object extraneous to the background. A layering approach (not yet introduced in our system) could help the algorithm to signal the bag as a stopped object.

We also show the background model computed by the HSV-SO algorithm and its change mask from previous frame. We would remark that the background model is represented by a neuronal map whose size is 9 times greater than that of the original image. In the reported figures they appear to have the same size only for an easier comparison. We can observe that the background model is a quite accurate (enlarged) representation of the real background. Small differences with the real background can be noticed only in few pixels near the column; these are due to a previous passage of the man in front of the column and the consequent only partial update of the corresponding background pixels.

Previous frame	Actual frame	Actual frame with moving object detection mask	Background model	Background model change mask from previous frame

Shadow suppression algorithms, such as the one described in [2], are readily inserted into HSV-SO background subtraction and update algorithm, having care of not updating the neuronal map for background pixels detected as shadows. An example is here reported:

Without shadow removal	With shadow removal

2. Sequence Walk1

Size: 384*288
Source: [3], filename: Walk1

We report two consecutive sequence frames (Previous and Actual frames) and the actual frame with corresponding moving object detection mask computed with HSV-SO algorithm. The man in the center of the room is perfectly detected, even though some parts of the man (such as the arm) tend to camouflage with the pavement and could have led to a partial detection. The group of persons lying close to the lower left side of the image (in the reflection on the pavement of the light coming through the windows) is partially detected. This is reasonable since such persons are barely distinguishable also for the human eye. Same observations hold for the person lying close to the plant in lower center side of the image.

We also show the background model computed by the HSV-SO algorithm and its change mask from previous frame. We can observe that the background model is a quite accurate (enlarged) representation of the real background. The change mask of the background model shows clearly that wide areas of constant intense white (in the reflection on the pavement of the light coming through the windows) are not updated from previous frame.

Previous frame	Actual frame	Actual frame with moving object detection mask	Background model	Background model change mask from previous frame

3. Wallflower Test Images

Size: 160*120
Source: [4], filenames: Camouflage, TimeOfDay, WavingTrees

The three image sequences represent some of the canonical problems for background subtraction highlighted in [4]. For each of them we show the first image of the sequence, the image in the sequence at which the comparison was done, the hand-segmented images of the foreground used as ground truth for comparison, and the foreground masks obtained with:

1) HSV-SO algorithm, described in [1];

2) Pfinder algorithm: the background model assumes that the intensity values of a pixel can be modeled by a Gaussian distribution [5];

3) VSAM algorithm: implements the approach proposed in [6], based on the integration of pixel analysis and region analysis modules to extract motion by a finite state machine. It is able to recognize when objects have stopped and to disambiguate overlapping objects;

4) CB algorithm: vector quantization is used to incrementally construct a codebook in order to generate a background/foreground model [7]. Results of the CB algorithm have been obtained using the software provided by the authors.

For all considered algorithms we experimented with different settings of adjustable parameters until the result seemed optimal over the entire sequence.

Finally, for each algorithm we report Recall and Precision measures, where

Recall = (number of true foreground pixels detected)/(number of foreground pixels in the ground truth),

Precision = (number of true foreground pixels detected)/(number of foreground pixels detected).

and we report also the F-measure (harmonic mean of Recall and Precision): Fmeasure = 2*(Recall*Precision)/(Recall+Precision).

	Camouflage	TimeOfDay	WavingTrees
First image of the sequence
Test Image
Description	Foreground covers monitor pattern	Light gradually brightened	Tree waving
Ideal foreground
Foreground with HSV-SO algorithm
	Recall=0.9094 Precision=0.9834 Fmeasure=0.9450	Recall=0.7637 Precision=0.8804 Fmeasure=0.8179	Recall=0.9858 Precision=0.9712 Fmeasure=0.9784
Foreground with Pfinder algorithm
	Recall=0.9553 Precision=0.9830 Fmeasure=0.9689	Recall=0.5021 Precision=0.9852 Fmeasure=0.6652	Recall=0.9922 Precision=0.6952 Fmeasure=0.8176
Foreground with VSAM algorithm
	Recall=0.9487 Precision=0.9796 Fmeasure=0.9639	Recall=0.4993 Precision=0.9878 Fmeasure=0.6633	Recall=0.9947 Precision=0.9552 Fmeasure=0.9745
Foreground with CB algorithm
	Recall=0.9700 Precision=0.9893 Fmeasure=0.9796	Recall=0.7198 Precision=0.9450 Fmeasure=0.8172	Recall=0.9727 Precision=0.9670 Fmeasure=0.9698