Development of Crowd Behavior Analysis Techniques

Abstract

Automation of behavior analysis of crowds is a challenging task due to the density variations, occlusions, context, and scene variability. This dissertation aims to address different challenges in image as well as videos of different crowd systems. The main goal is to help quantise the crowd behavior, segment salient crowd groups and detect false crowds in forged images using spatial and temporal cues derived from the crowd itself. An improved crowd coherent collectiveness descriptor for analyzing and measuring the collective motion of crowds is proposed. The descriptor is estimated based on velocities which can result in zero collectiveness for static crowds. To overcome this issue a moving weighted average concept is applied for computation of horizontal and vertical velocities. A voting based scheme is used to cluster the crowd. It takes the information of clustering from previous frames and based on the maximum voting results cluster the crowd at current time. The proposed scheme is useful for density estimation and behavior analysis of different crowds and its application for creatures (sea-birds, fish, and others etc.) is also explored. Visual and quantitative analysis verifies the significance of the proposed scheme. To quantify the crowd behavior we use different descriptors. The existing descriptors are contextual and generally provide information about crowd density categorized as low, medium or high. However, other properties of the crowd like speed, direction, shape and merging probabilities (of different crowds at group level) are also important for crowd analysis. In the proposed technique, crowd descriptors are introduced which are robust against different outliers and different densities of the crowd. Simulations on various datasets show the applicability of proposed descriptors. Crowd flow and dominant motion detection technique (locally between the clusters) are combined to detect crowd saliency in this research. For each cluster, the saliency map is computed using motion and contrast cues. The dominant motion is detected using k means clustering. The proposed technique is able to segment the crowd cluster with maximum motion and can even detect the straight-line motions. The proposed scheme can identify salient crowd groups. The applicability of the proposed scheme has been demonstrated for microscopic medical data. The motion of the immune system makes it more salient than the surrounding cells. Thus exploiting this idea a spatio-temporal technique is proposed. In which temporal saliency is computed using extended Lucas Kanade and coherent clustering while spatial saliency is computed on feature extraction. Simulations on different datasets show the effectiveness/ applicability of the proposed technique. The authenticity and reliability of digital images has become one of the major coni cerns recently due to the ease in manipulating and modifying these images. Similar manipulation in crowded images give rise to false crowd, where a person or group of persons is copied and pasted in the same image. Thus, the detection of such false crowds is the focus of current research. In this research, false crowd detection in forged images is carried out using a modified and improved PatchMatch algorithm which can even detect multiple copies of the same instance. To separate humans from non-human objects a human detection algorithm is used in the post-processing phase. A benchmark database consisting of false crowd images has also been developed. Experimental results confirm that the technique is capable of detecting the false crowds successfully and is even robust for multiple cloning problem. We experimentally demonstrate the effectiveness and robustness of proposed algorithms by quantifying crowds at the group level, segment salient crowd groups and detecting false crowd groups.