Vision-based Algorithm Of Moving Object Detection And Tracking In Complex Scenes

Intelligent video surveillance is the system which can analyze and understand the video signals without the human intervention. In order to make the human liberate from the work, the system should have the same intelligence as human. In the aerospace, the space capsule is a complex and narrow space, astronauts and the equipment in the capsule are interdependent. In microgravity environment, there are many uncertain factors, which need the astronauts make the right and timely judgment.Therefore, studying the movement strategy and behavior of the astronauts,especially in the early research stages of the space station is very meaningful.We have to do in-depth study about the astronauts operate and maintain the equipment; their bodies comfort and fatigue, whether hurt or not; the workload and if the collaboration is smooth. These issues not only relate to the safety of the astronauts, but also optimize the space capsule design.Therefore, in order to solve these problems, we set up a video monitoring system and launch on collecting motion parameters of the astronauts in space capsules. It includes the rapid detection, pose estimation and trajectory tracking of the astronauts. Firstly we propose the approach that based on coarse-to-fine fast object detection based on hierarchical model; then multi-view segmentation and graph model for pose estimation method; and level set segmentation and object tracking algorithm with context; in the end we use the CRF learning model for tracking framework. These approaches can provide the reliable basis for the collecting of the astronauts' motion parameters, and the ground simulation experiments confirm the reliability of the experiment.The dissertation includes following sections:Firstly, for deformable objects detection in complex scenes, we propose a fast target detection method which based on hierarchical model with coarse-to-fine reasoning. We established a new hierarchical model, the first feature level introducing conditional random field model to fuse local salient features to improve the accuracy of the candidate feature area. Parts layer,which get from the regular segmentation and root parts recursive. A coarse-to-fine inference algorithm is matching from the low-resolution parts.The experiments show that our detection method is effectively to the detection speed, and it has high detection accuracy.Secondly, for the goal of pose estimation in complex scenes, we propose the method which based on multi-view segmentation and graph model to achieve pose estimation. First, using the graph structure model (PS)to divide the human body into several parts, and then combine the various components to represent the various human pose. Meantime we combined the image features, edge feature and regional feature as adjuvant. We combine multi-view robust segmentation (maximum a posteriori Markov random field (MAP-MRF)) to achieve the multi-pose estimation. In addition,we introduce a novel shape priori combine with the original pose estimate to segment the objects in the image. Thus we can greatly reduce the blur. The experiments show that our method can effectively complete the pose estimation, in the case of occlusion and interactive we still can get the robust pose estimation.Thirdly, for the case in complex scenes, there are multi-articulation targets, how to obtain the feature effectively and segment the target accurately while with the multi-target tracking. We propose that fusion context information and level set multi-region segmentation for tracking.First of all: In the segmentation stage, we propose two improvements.Second: In the tracking part, we propose a method to solve the interference of similar appearance objects and objects leave the field of view by use of context information and establish two auxiliary items. They played an important role in checking the real objects. A large number of experiments in the real-world, we found that use the method that we proposed makes accuracy of tracking improved.Fourthly, research the affinities and dependencies, use CRF modeling multi-target tracking. We use CRF model to evaluate the affinities and dependencies between the targets during the tracking. Experiments show that the method is effective for tracking multi-targets, and robust to the occlusion.Finally, based on the ideas of the subject, we build a simulation system of experiment platform to test the proposed algorithms in the actual scene of the school. Experimental results demonstrate the feasibility and correctness of the algorithms. And in the future, through the study of target tracking and detection methods we found that in this area there are many points need further study, and the accuracy of detection and timeliness of tracking can be improved. We proposed that in the future the improvements in algorithm optimization itself; and followed in the processing of the target occlusion and deformation as well as the tracking in the complexity scenes, more innovative ways and means will proposed on both hardware and software to improve the robust about tracking.