Research Of Algorithm For Pedestrians Crowd Behavior Analysis And Recognition

At present, the pedestrian crowd behavior Analysis and recognition research mainly focused on the Multi-Target Detection and Tracking(MTDT), Trajectories Analysis(TA), Dense Regions Analysis of Pedestrians(DRAP), etc. This thesis is based on P. F. Felzenszwalb proposed pedestrian detection algorithm named Mixtures of multiscale Deformable Part Models(MDPM) and Our proposed new pedestrian detection algorithm. This thesis mainly improves the multi-target tracking and analysis method. This thesis proposes a new feature extract method for multiple target tracking. This method can significantly improve the pedestrian tracking performance. And we do TA and DRAP based on the tracking results. The result is helpful to square, building, subway, airport. We can do crowd behavior analysis, forecasting, early warning in these scenarios. The main research content is as follows:1. Put forward the LBP texture feature and Histograms of Oriented Gradients(HOG) fusion feature HOLB extraction method. And then put forward a new pedestrian detection method with Adaptive Boosting algorithm based on HOLB feature(HLAB).2. Put forward the LBP texture Similarity feature and Distance measurement fusion feature LSED extraction method. Put forward a new Model(LMM) for group pedestrian tracking based on the LSED feature and MCF. The model treats two adjacent frames as a unit, pedestrians in these two frames as the nodes. Calculate the cover rate of two pedestrians between these two frames, if the cover rate greater than the threshold, then connect the corresponding node in the model, with corresponding LSED feature as the Cost. Set up the Capacity of the edges in the model, then the model of the two adjacent frames is formed. Take two adjacent frames in the video to form two frames LMM model in turn. Experiment proved that this model has a low IDS times in group pedestrian tracking and the results contain the complete detection results.3. By using the pedestrian tracking results of LMM model, can further analysis the pedestrian crowd behavior, mainly includes TA and the DRAP. TA mainly analyzes the whole distribution of track fragments and the causes of fragments. DRAP statistics pedestrian numbers within a specified radius around each pedestrian as the density at this point, form the DRAP distribution. Experiments prove that TA can directly reflect the existing problems of the tracking results, and DRAP can intuitive reflect dense area of pedestrian.Finally this thesis puts forward the LMM, TA, DRAP combined new algorithm LMTDR based on HLAB, used for crowd behavior analysis and recognition of pedestrians.