A COMBINED HISTOGRAM OF ORIENTED GRADIENTS AND COMPLETED LOCAL BINARY PATTERN METHODS FOR PEOPLE COUNTING IN A DENSE CROWD SCENARIO

Estimating the number of people in a dense crowd scenario is one of the most
interesting subjects in visual surveillance system application. It is extremely important
in controlling and monitoring the crowd for safety control and urban planning.
However, estimating the number of people in any dense crowd situation is not an easy
task. This problem mostly arises due to some false positive and false negative and it
affects the performance of system on detection rate. Therefore in this thesis, an
innovative method for people counting in dense crowd scenario is proposed. This
method used a collaborative Histogram of Oriented Gradients (HOG) and Completed
Local Binary Pattern (CLBP) based on people detection algorithm to detect headshoulder
region. Head-shoulder region is used as features to detect people against the
false positive and false negative issue. HOG and CLBP descriptors are utilized to
extract the edge contour and texture features of head-shoulder region, respectively.
The two features are then fused together to generate a cumulative feature vectors.
Support Vector Machine (SVM) is used to perform classification of the fusion
features to people from a mixture of objects. The results show that the detection rate
of the proposed method HOG-CLBP, on Recall value and Accuracy, achieves better
performance compared to the current method for dense crowd scenario.