Results obtained from the SMG method

We find considerable improvement combining background subtraction with gradient based activity detection method. The results reflect the benefits of having a slow adapting background model along with a fast and strictly motion based region of activity generator. GMM based background subtraction image incorporates a lots of errors induced due to many factors varying to switching on/off of light sources to noise presence in the frame due to highly illuminated scene capture by the sensor. The GMM based model adapts slowly to the changes occuring to the background. So when a light is switched on/off, the region affected by this change gets classified as foreground. But the fast SMG model rejects that region of change immediately unless that light source is repetitive in natural or has motion associated with it. Since SMG is a pure motion based region of attention predictor so when the motion is minimal, it switches to buffered region of activity image for help. SMG based model doesnt affect the adaptive nature of the GMM. Its a tool to get rid of the unnecessary errors incorporated in the background subtraction results from the GMM method.

The results presented in the Figure 3.3, show the foreground objects with very low granular noise compared to previously mentioned algorithms. The foreground objects are registered or detected as blobs and the number of false positives has decreased significantly. Figure 3.2, shows the region of activity or RA images corresponding to each of the output image obtained in Figure 3.3. As we can see, region of activity images do not provide us the complete information about the activity region but its sufficient enough to extract the foreground belonging to that portion of the image. Figure 3.4 shows the implementation of above method on the data used in the Genetic Algorithm section and the results obtained over this set of data are also very good. In the Figure 3.5, we have a foreground object in a highly illuminated scene. The foreground object switches off the light in the frame-1714, and we can see the affect of it in the very next output frame 1715, but the SMG method helps it recover quick enough as can be seen in the output frame 1716. Figure 3.5 further elaborates the extend of usefulness of the SMG module. As we can see due to fast adaptive nature of the SMG module, it rejects the regions affected by illumination which are classified as foreground. As described by Jwu-Sheng [12] we need a long term model which is the GMM based background model because of its adaptive nature but along with it we need a short term model which takes care of short term tendencies of the scene, which is the proposed SMG model.

Figure 3.2: Region of Activity frames for data used in Figure 2.1

Figure 3.3: Results of background subtraction for frames in Figure 2.1 using the proposed SMG method. Comparing figure (d) and (e) with (g) and (h), we can see vast improvement in reducing false positives.

Figure 3.4: Background subtraction obtained from the SMG method using the Genetic Algorithm section data

Figure 3.5: Other results using SMG method where a person is switching off lights

Figure 3.6: Actual Background subtraction images from optimized GMM method without using SMG module. Note the high rise in false positives once the light is switched off in frame 1714.