| Recently,multi-object tracking is considered as a research hotspot in computer vision.It is commonly seen as the basis for advanced visual tasks in many areas,such as behavior recognition,and so on.It also involves many interdisciplinary subjects such as probability statistics,machine learning,and image processing.Multi-object tracking can be widely used in intelligent video surveillance,robotics and military industries and many other fields.As the main research direction in the field of multi-object tracking,the difficulty of pedestrian multi-object tracking is to build a more complete strategy or system to effectively deal with feature extraction,occlusion handing and data association problems in complex scenes.Among them,the design of the feature model and the processing of the occlusion are crucial to the tracking effect.This thesis mainly studies the multi-object tracking algorithm based on feature consistency.Our thesis begins with a summary of the classic algorithms in this field in recent years and briefly outlines the basic theory of multi-object tracking.The algorithm adopts an appearance model based on feature consistency theory and an occlusion handing model based on motion confidence,and constructs a pedestrian multi-object tracking system.Based on previous research,the main work of this thesis is as follows:(1)The occlusion problem is studied and a new occlusion handing model based on motion confidence is constructed.Experiments show that the model is robust to target occlusion in complex scenes.(2)For pedestrian targets,color features,texture features,motion and size information based motion features,combined scenario information are used to construct a feature consistency model.(3)Combined with the feature consistency model and occlusion handing model,the trajectory classification strategy after data association is studied and optimized.The simulation of our algorithms was completed,and all parameters were adjusted for the tracking scenario.For the real scene of video surveillance and the datasets of TUD、PETS、ADL、ETH,tests have been completed,which achieves a comparable tracking effect. |
