Research On Data-driven Crowd Analysis And Multi-scale Simulation

With the advent of big data and the age of intelligence,techniques based on data-driven,which usually function by mining knowledge and identifying patterns from real data,are becoming one of the most effective ways to solve problems in both academic and industrial world.Among them,the data-driven crowd scene monitoring,analysis,behavioral prediction and motion synthesis are also becoming hot topics in the field of crowd simulation.Since the crowd is a collection of multi-agents,considering the diversity of individual,environment and the complexity of organism,implementing data-driven based techniques rather than drawing up limited rules for crowd simulation makes the research direction shift from the field of human mechanics,psychology and physics to the field of pattern recognition and machine learning.However,due to the complexity and flexibility of crowd data,it still faces many challenges to completely rely on data for macroscopic scene modeling,mid-view collision avoidance,as well as microscopic motion synthesis in simulating crowd.Supported by data-driven based techniques,this dissertation has carried out a series of work on the analysis and simulation of crowd in the direction of macroscopic,mid-view,and microscopic point of view:A new pedestrian detection method based on feature weighted Support Vector Machine(SVM)is proposed to detect and monitor pedestrians in video and image.The core idea of this method is to use the Genetic Algorithm(GA)introduced with Metropolis criterion(MGA)to set weights of data features in SVM.Specifically,in MGA,the Metropolis criterion is adopted into GA for dynamical parents' selection,which makes the algorithm get a stronger ability to jump out of local minimum as well as achieve convergence.Experiments show that by applying this algorithm into the feature weight learning scheme in SVM,in the field of pedestrian detection applications,the detector can achieve a more accurate detection result.A novel crowd trajectory based macroscopic scene modeling and simulation framework is proposed to further expose the unique attributes of crowd data in real scenarios,and finally positively guides the simulated crowd towards real scene.The framework first analyzes scene information based on the information of historical trajectory data to determine the motion pattern.Then,considering each cluster of motion trajectories,the velocity field is calculated separately,and anomaly removal is performed.Finally,the macroscopic scene modeling and path planning are guided by the purified velocity field to match the space-time properties of real trajectories.Furthermore,in the experiment,we find and confirm the following three spatiotemporal statistical properties of pedestrians inside each motion pattern:(1).The distribution of path length obeys the power law.(2).Pedestrians' speeds follow a Gaussian distribution.(3).Pedestrians tend to maintain a lower speed in entrance/exit and turning areas while a higher one in the middle of a given path.A crowd simulation method based on multi-hidden layer neural network is proposed to implement the local collision avoidance operation in mid-view.The core idea of this method is to train a multi-hidden layer neural network to fit the state-action data pairs in the crowd,and finally get a network model embedded with pedestrian motion decision,thus guiding the agent to perform local collision avoidance.The proposed method firstly meets the need of traditional data-driven methods for authenticity of crowd simulation.At the same time,it is superior to the traditional search-based data-driven crowd simulation methods in the number of collisions and operating efficiency.What's more,the neural network model can learn some non-existent effective motion decisions inside the training data,thus further illustrating the effectiveness of the proposed simulation method in guiding agents to perform local collision avoidance.Deep Neural Network model is further explored.A Consistency Term(CT)penalty function is proposed to impro've the accuracy of the deep neural network as well as the quality of samples generated through the generative model.To further restrict the Lipschitz continuity of the discriminator,we devise a regularization over a pair of data points drawn near the manifold following the most basic definition of the 1-Lipschitz continuity.In particular,we perturb each real data point x twice and use a Lipschitz constant to bound the difference between the discriminator's responses to the perturbed data points x',x".As a result,it gives rise to not only better photo-realistic samples than the previou.s methods but also state-of-the-art semi-supervised learning results.Finally,we implement this CT loss into local collision avoidance and achieve a smaller testing error compared with the original multi hidden layer neural network.A Hybrid Motion Graph(HMG)framework is proposed to achieve a more flexible synthesis of pedestrian motion on the microscopic level.The fr-amework draws on the idea of Motion Fields for the flexible synthesis and control of motion inside each motion class.In the meanwhile,the framework implements Motion Graphics technique among motion classes to generate fast and flexible transitions.Specifically,in constructing the HMG,firstly,the motion template of each class is automatically derived from the training motions for capturing the general spatio-temporal characteristics of an entire motion class.Then,a typical motion fields is constructed for each class including each type of motion along with its corresponding motion template.The next step is to interpolate motion templates with space-time constraints and then generate a smooth transition between each motion class pairs.Specifically,the motion fields of each class are integrated into the global structure control of the motion graph,thereby establishing smooth.transitions between templates and motions inside each motion fields.At last,we summarize this dissertation and discuss some prospected work.