Improvement And System Implementation Of Pedestrian Detection Algorithm Based On DPM And XCF

Pedestrian detection technology is one of the most focused and difficult research in the field of computer vision.It has a great application prospect in video surveillance,intelligent transportation and other fields.But the application of pedestrian detection is restricted by two points: On one hand,pedestrian targets are non-rigid targets,which can be deformed easily and occluded by objects in the environment,leading to lower detection accuracy.On the other hand,the real-time ability of most pedestrian detection algorithms is poor,which limits the application field.In this thesis,two popular pedestrian detection algorithms are studied and some modified algorithms are proposed based on these two algorithms.Deformable part model is a successful algorithm in pedestrian detection,which uses deformable multiple part models to effectively deal with the problem of partial occlusion and deformation.The pedestrian detection algorithm based on channel features is also a research hotspot in recent years,which exploits the characteristic of natural feature fusion to excavate the advantages of various types of features.In addition,a fast feature pyramid construction method is also used to achieve higher detection speed.Author of this thesis studies the advantages and disadvantages of above algorithms then proposes some improvements and innovations including following aspects:(1)To solve the problem of insufficient accuracy of pedestrian detection algorithm,a fusion method based on the detection score of various pedestrian detection algorithms is proposed.This thesis studies the information fusion theory based on matching level as well as its various application situations and uses this theory in the pedestrian detection algorithm.Simulation results show that the miss rate and mistake rate are significantly reduced.(2)Since channel feature based detection algorithms have the advantage of speed and the part model can solve occlusion and deformation problems,this thesis utilizes a simple part model based on channel feature,which can improve the detection accuracy while maintaining high detection speed through choose proper part model according to the experiments.(3)A variety of feature selection methods are used to analyze the channel features of the ACF algorithm,and found that there are many invalid and redundancy features in ACF features channels.The results indicate that the performance of the algorithm could increase a certain extent,by restrain the role of these features in classifier during the training session.Finally,a pedestrian detection system is constructed.For different monitoring environments,the pedestrian detection scheme with better speed or accuracy is adaptively chosen.At last,the valid information of pedestrians is extracted and saved in the database for quick retrieval.