Detection and Classification of Moving Objects for an Automated Surveillance System

Automated surveillance system has been the subject of much research recently. A completely automated system means a computer will perform the entire task from low level detection to higher level motion analysis. Since conventional system practically using human power to monitor and did not applicable for a long hour monitoring, thus automated system had been created to replace the conventional system. This thesis focuses on a method to detect and classify a moving object that pass through the surveillance area boundary. Moving object is detected by using combination of two frame differencing and adaptive image averaging with selectivity. Technically, this method estimate the motion area before updates the background by taking a weighted average of non-motion area of the current background altogether with non-motion area of the current frame of the video sequence. This step had created a focus of attention for higher level processing and it helps to decrease computation time considerably. The output of a motion–based detector is essentially a collection of foreground that might correspond to the moving objects. But usually the output image produced from this iii process contaminated with noise and shadow. As a solution, morphological operation has been employed as an approach to remove noise from the foreground object. Mutual shadow that exists with the object had been abolished by combining chromatic colour values with lightness variable. Then, standardized moment invariant is employed to extract the features for each moving blobs. To recognize these blobs, the calculated moment values are fed to a support vector machine module that is equipped with trained extracted moment values for human and vehicle silhouettes. The system operates on colour video imagery from a stationary camera. It can handle object detection in outdoor environments and under changing illumination conditions. The applied post processing module capable to remove noise and shadow from the detected objects with less than 1% of error. Finally, classification algorithm that makes use of the extracted moment values from the detected objects successfully categorize objects into pre-defined classes of human and vehicle with 89.08% of accuracy. All the methods have been tested on video data and the experimental results have demonstrated a fast and robust system

Preview PDF 600440_fk_2006_87_abstrak_je__dh_pdf_.pdf Download (114kB)

Actions (login required)

View Item