High Density Crowd Segmentation And Behavior Recognition

Intelligent monitoring technology can make computers have the ability to analyze results independently by using computer vision, digital image processing, pattern recognition and other related domain knowledge, timely recognizing different kinds of behaviors in the true scenes. The study found that emergencies are more likely to happen in the high density crowd. Due to the inherent complexity of the high density crowd, finding an effective method to analyze and understanding the behavior has become one of the focuses of scholars.Combined with the latest computer vision research theory about the high density crowd behavior analysis, we implement in-depth research for some relative specific problems, such as crowded scenes representation method, behavior segmentation, feature extraction and recognition technology. The exciting parts of our article are four aspects:1. We introduce a novel representation method in detail based on streaklines to describe high density crowd scenes. For crowded scenes, severe spatial changes and artificial time lag appear in behaviors when we use optical flow method and the pathlines to describe scenes. So we introduce streaklines concept based on Lagrange method in order to remedy the problems mentioned above..2. This paper segment every frame into regions based on the similarity of the neighboring streaklines and streak flow. In our paper, the streaklines similarity is mainly manifested in two aspects:(a) Unit projection of motion vectors angle similarity measure is used to distinguish different motion; (b)Distance measure, that is, particles relative motion trend change can reflect the consistency of motion. The similarity of the flow field is defined as the velocity angle of streak flow at a certain moment t in a fixed point p. Using the weighted similarity measure value and watershed segmentation algorithm, we segment crowed scenes into different region according to the different behaviors. In addition, we also introduce Lyapunov coefficient on processing the over segmentation phenomenon.3. For crowded scenes, target detection and tracking are impractical and not necessary. So we propose a holistic strategy to make characteristics extraction and scenes description, namely the space-time cube features for Micro-behavioral semantic based on dynamic system analysis. Firstly, by using the improved flow field, we analyze the stability of the dynamic system, and set up five kinds Micro-behavioral flow model, respectively named bottleneck, fountainhead, blocking, lane and ring/arch; Secondly, Jacobian matrix of effective particles in block area is computed, eigenvalues in different states will be used to judge particles which Micro-behavioral flow belonging to, and the proportion of different flow’s particles will be as descriptors for each block to establish histogram. Lastly, taking full account of the spatial and temporal characteristics of crowed behaviors, our algorithm constructs space-time cube characteristic for Micro-behavior semantic, obtaining the expression of high density crowd behaviors information. As a kind of intermediate semantic features, Micro-behavior semantic take full advantage of the underlying local motion feature information, and make combination with space-time cube high-level information characteristics, solving the semantic gap problem.4. Due to traditional sparse representation method unstable characteristics endemic structure in crowd behavior detection, we propose a new crowded behavior analyze method based on Locally linear embedding (LLE) and sparse representation. The local linear embedding regularization can effectively preserve the local manifold structure of test sample, and improve sample discriminant ability. Experimental results show that the local linear embedding algorithm of sparse representation can effectively improve sample discriminant ability and get good experiment effect in the high density crowd behavior detection.We prove our proposed segmentation algorithm and behavior recognition technology using the UMN and UCSD1 public video database. The experiment results show that our algorithm can effectively segment and recognize crowded behaviors.