Background Subtraction in Highly Illuminated Indoor Environment

Tarun Baloch

Supervisor: Prof. Amitabha Mukherjee

Abstract

In this work we present a new method for attaining robust background subtraction in Indoor Environment with variable illumination conditions, reflecting surfaces, shadows, flickering of light sources and monitors, sensor noise etc. For the system to be adaptive and realtime, each pixel is modelled as a mixture of gaussian(GMM). GMM based background subtraction was able to recognize the foreground object but the noise incorporated in the background subtraction results due to the above mentioned factors, made it a very difficult task to extract the foreground object. We will also talk about few of the other popular methods in background modelling which failed to give us good results in varying illumination conditions. To get rid of this noise and improve the background subtraction, we propose a new Short Memory Gradient(SMG) based background subtraction method. We use the edge information obtained from the frame subtraction of the present frame from the previously encountered frame and an approximately build background frame. The edge information obtained from the above is combined together and is used to reconstruct a probable Region of Activity(RA). This probable region is combined with the background subtraction result obtained from GMM to reject the false positives and extract the actual foreground object. The reason for using frame subtraction is to incorporate motion as a verifying feature alongside the color.

We also present a digital video recorder(DVR) system which captures images from an IP camera and stores them in a structured directory for a certain time period. The system starts overwriting the previously stored data once the cycle of that time period is completed. An Online Query Interface is developed for users to easily browse through the stored data. Queries are necessarily based on time intervals, where the user is presented with the frames recorded during that period of time. This process could become tedious when the query time intervals are big and further more irritating when there is no activity happening during that time period. In those cases, the user is provided with an option to switch on the activity bit which would provide the user with only those frames where some foreground activity is registered in the scene.

Contents

Thesis Report

References

Code