논문 과제

A Real time unoccupied object detection through Background Modeling
*Muhammad Adeel Altaf
School of Electronics Engineering
Kyungpook National University
Daegu, South Korea
adeel_farukh3@hotmail.com


Abstract— A technique is required to detect the unoccupied objects from surveillance video. The method utilizes a Background modeling algorithm and identifies any static object lying unoccupied for a specific period of time. At first, with the help of Background Subtraction, foreground objects are extracted. Background Modeling is used through running average-based method. In the second step, static objects of foreground objects of consecutive frames are detected. Experimental results identify that proposed algorithm is efficient for real time video surveillance which is tested on ABODA datasets. The approach uses a FIFO queue for each pixel in a video frame as an essential element with applying processing method. The pixels of moving objects are visible but they cannot store in the queue. In the last, static objects can also be further classified into human and non-human objects with the help of edge-based object recognition approach. The technique successfully detects abandoned objects of different sizes present in the video.
Keywords—Abandoned object detection, Video surveillance, Background Subtraction

I. INTRODUCTION
It is becoming a global problem that suspicious activities are emerging everywhere. It is needed to identify those things such as if someone comes in gathered areas like railway stations, bus stations or even airports and leaves a bag in public places, it may be contained some explosive material. To detect the abandoned luggage, an intelligent system is required which can detect the unoccupied objects from a gathering environment as well as it can report to the relevant security staff to avoid the incidents with the help of a sound producing alarm.
It is always considered, especially in rural area, as a quite difficult task to watch the activities all the time in public areas with the help of Closed-Circuit Television (CCTV) cameras. If somebody leaves its luggage anywhere, it is risky to avoid those things for a long time unoccupied. However, the current surveillance system is a man-powered, where a security guard must have to constantly watch the cameras all the time, throughout the area to find anything suspicious activity. It looks easy but it is not to keep your eyes on the screen for hours. And most probably, he may miss any criminal activity by looking here and there.
Several alternatives are presented to solve this problem. High resolution overhead cameras are used for surveillance in significant locations such as official places or private establishments, to detect such unattended objects of even small size. Although, these cameras are costly, and even with their resolution, there is no guarantee that the security personnel will spot the object. In some cases, security officers often patrol the area under surveillance, to maintain vigilance over any such object. This is commonly seen in railway stations and areas of high-security zones. However, this tactic requires excessive

manpower, especially if the area under surveillance is huge. Also, if there happens to be a bomb concealed in an abandoned object, this raises the risk of physical harm to the officers who would examine this object.
It is also necessary to educate the people that as they notice any unoccupied object anywhere, it’s their responsibility to raise and discuss the issue with the officials immediately. Through advertisements on the TV channels or the social media, humans can be motivated to identify such incidents for their safety. As a matter of fact, people ignore such type of instructions and mind their own business. In fact, if there is a real case that someone finds a bag, he himself feels awkward to inform the security officials. So, in these circumstances, it is needed to build an autonomous system which acts intelligently under these conditions.
The proposed method takes a video stream as an input and models a static frame as a reference in which other subjects are supposed as ordinarily present in it. First, it does background subtraction to find out the foreground objects. Then it updates the background in which the object that has changed its original position or any new object comes into its coverage are of region that has been abandoned. At last, it compares the new background with the original frame which contains the unoccupied object in its new window.

II. PROPOSED FRAMEWORK
Several methods have been previously proposed for the detection of abandoned objects. In [1], Miguel et al. fused three features based on shape and color information to detect unattended or stolen objects. Jing Chang [2] utilizes selective tracking to determine whether the owner of the so detected abandoned luggage is in proximity by detecting “skin color information” and body contours. However, it is difficult to determine in crowded situations where people are rapidly moving towards and away from the luggage. Also, detecting such object of a person is subject to lighting changes and occlusions due to a crowded environment. Most of the proposed methods, such as the one proposed by M. Bhargava [15] search for an owner of the luggage in the midst of the crowd, which requires the system to have a lot of memory at its disposal. Fan et al. [13] proposed a system to detect large abandoned object with low false positive rates. Background Subtraction [7] is also commonly utilized, but it is a practical only if the background so acquired updates itself with lighting changes over time, to prevent noise contours. Gaussian Mixture Model (GMM) is commonly utilized to achieve background updating [12] [3] [10]. Although GMM is an efficient method for background modeling but it is memory intensive. So, it is needed to present a memory efficient algorithm to achieve a result that is comparable to that of GMM.

Fig. 1: A static frame- I without object and II with object

그림 1: 정적 프레임- 물체가 없는 I 및 물체가 있는 II

III. ALGORITHM
A. Defining Reference Static Objects Frame
This approach assumes during the preprocessing step where it considers that there are no any unoccupied objects in the ‘n’ frames of the video. So, in order to make an initial static background from the input video, it contains only those objects which are at rest position. The composition of static objects is stool, fire extinguisher, electric board, alarm system and so on. It is done by using the average value of each pixel in the first ‘n’ frames making an image. It is computed just like the design flow of the algorithm within seconds and is capable to store its static background in its memory as well.


B. Upgrading Current Static Framework
As the preprocessing is done with the help of static objects in the background that are usually present in such type of videos, it is definitely needed to update the previously defined computations and make a model of an upgraded static background. So, this is done after the desired time interval by incorporating with the new incoming static objects in the input video. Moreover, if a camera detects a new object within its defined range and it does not change its position with certain time period, it can be seen in the upgraded static background. In the situation, it will still work even if the object is not cleared due to the low resolution of the video. Although, the speedily moving objects have a low impact while upgrading the static background, so this updating background is unaffected by object occlusion due to the crowds. It is also helpful in detecting the objects in the presence of temporal as well as spatial occlusion.


C. Frames Comparison
It can be considered that if any object is not detected in the video feed for an extended time period, its presence can be visible in the upgraded static frame. This updated frame is compared with the previous ones without having the object, such comparison is performed by calculating a difference between both frames.

D. Classify into human and Non-human objects
By using edge-based object recognition approach, it can be categorized the objects into human and non-human static object. This algorithm basically uses edge information of an object into a human edge-based sample [14]. After the matching-template generated of static object with the pre-defined sample of human edges, Template matching is performed to generate a score. So, this generated score classifies the static objects into human and non-human objects. In this regard, Edge-based Object Feature and Template matching score generation steps are performed to produce the matching score between the predefined human-based template and the detected static object.

Fig. 2. Pixel-based Background Modeling Object detection

V. MATHEMATICAL PRESENTATION

Consider the nth frame of the video as In, such that In (i, j) represents the value of pixel present at ith row and jth column of nth frame. For each such pixel value a queue Q (i, j) of size N, a sum of pixel values S (i, j) and average of pixel values A (i, j) is maintained over the incoming frames. The total number of frames N to be considered for modeling a static background so, it can be selected by user as per the requirement. It is directly related to the measure of time after which an object will be declared abandoned. Starting from the first frame, n = 1: