Crowd detection is a highly focused area for law enforcement, urban engineering andtraffic management. Public places such as shopping centers and airports are monitored usingclosed circuit television in order to ensure normal operating conditions. Automated analysis ofcrowd activities using surveillance videos is an important issue for communal security duringviolence, strike, heavy gathering allows detection of dangerous crowds and where they areheaded. In case of surveillance, group behavior modeling and crowd disaster prevention peopledetection and tracking in crowd is a crucial component for wide range of application. Due toheavy occlusions, view variations and varying density of people as well as the ambiguousappearance of body the reliable person detection and tracking in crowd becomes a challengingtask. Computer vision based crowd analysis algorithm can be divided into some groups; peoplecounting, people tracking and crowd behavior analysis, movement analysis.Person detection and tracking in crowd is a challenging task.
Individual object detectionhas been improved significantly in recent times but the crowd detection and tracking containssome challenges. The head density of one person could be similar to another person density.The density of pedestrians significantly impacts their appearance in a video. For instance, in thevideos with high density of crowds, people often occlude each other and usually few parts of thebody of each individual are visible. On the other hand, the full body or a significant portion of thebody of each pedestrian is visible in videos with low crowd-density.
These different appearancecharacteristics require tracking methods which suite the density of the crowd 1.This research proposed a system that detect the head region and based on this headregion that can detect people from crowd. The more accurate head detection can lead a goodresult for detecting a person in crowd domain. This research focused on gradient feature basedimage analysis and found a good accuracy rate of head detection described based on belowFigure 1 and Figure 2 as sample.Figure 1. Sample Input Image to Detect Crowd Figure 2. Detected Crowd ResultConcentrated on image gradient based people detection.
Image gradient basicallycontains the directional changes information. This information can be used to track differentobjects or regions as well as boundary shape, getting a rough idea of an object location andother information. By analyzing the regions an assumption of item can be found that it is humanor not in another word we can say that this step is important part selection or interest pointdetection. Natural images contain a lot of changes in orientation. So the number of importantpart may be large and huge as it is counting based on orientation information.
There is sometypes of methods is needed to reduce this important part such as Adaboost and others. Weapplied different feature extraction technique to detect human on that region or from crowdplace. We analyzed with HOG, SIFT and SURF feature.
We used HOG and SIFT combinedfeature to test the result.Assume a section that is a strong candidate for head region means an interest pointthat’s may be head or not, is compared with trained support vector machine. Applying manualannotation technique we have prepared two classes of data, one is positive dataset another isnegative dataset. During dataset preparation we have developed dynamic patch selection andits size.
Supervised SVM is used to train with two dataset. All the candidate regions are testedwith SVM. This test said which one is head or not. We got a marked output that processed withproposed method.Next sections are organized as follow. Details of implementation in section three,preparing dataset with manual annotation in section four, experimental results are shown andcompared with different methods and the next section contains the conclusion.