Image/Video Motion Estimation And Analysis

With the rapid development of multimedia and sensing technologies,image and video,as the main channel of vision recording,play an increasingly important role in our daily life.In recent years,there are a large amount of image and video data booming in all kinds of aspects.In these visual data,motion information takes an important part in the visual data generation and content recording,covering all the domains of image and video processing,ranging from low semantic level like motion trajectory and high semantic level like motion understanding.Specifically,in the image generation,camera motion makes image blurry,and increases the entropy of image,in order to recover the latent image,we always need to estimate the camera motion accurately.In image content recording,object motion may cause image partial motion blur,this motion blur provides more information as blur captures a series of motion.Video can be treated as the temporal stacking of image sequences,similarly,camera motion may lead to the blurry video and cause video visually unstable and motion in the video content records the dynamic environment and agents,which is the main motivation of the video capturing.In this dissertation,we focus on estimating and analysing the motion of image and video,specifically including:in the image generation,camera motion modeling and representation,camera motion estimation and latent image recovery;in the video content,hierarchically human motion characterization and fast human motion detection.The main contributions in this dissertation are:1.Deeply analyzing the camera motion in image generation,we propose a novel separable kernel and its individual optimization model.The separable kernel can reveal the inherent camera imaging and provide a new viewpoint of image deblur-ring.To demonstrate its advantage,we propose a trajectory-based optimization algorithm for deblurring.For many cases,where current deblurring approaches fall into the local minimum,excellent deblurred results and correct blur kernels can be obtained by individually optimizing the kernel trajectories.2.Based on the observation that highlight regions can faithfully record the camera motion,we make the challenging night blurry image deblurring problem possible by deducing the kernel information from highlight regions.We propose a novel scheme to gracefully combine the deduced kernel from highlight regions and the non-highlight regions for accurate kernel estimation.Besides,we propose a novel function-form kernel representation and energy minimizing algorithm to automat-ically assign the deduced kernel to different regions for non-uniform deblurring.3.In the video content generation,we focus on the semantic understanding of motion information:fall detection as the pointcut,to satisfy the real-time motion analysis,real-time fall detection,we treat the motion information in a scalable way and build a boosted cascade for fall detection.The cascade contains the very different complexity of features which is quite different from tradition framework only containing a single type of feature.The cascaded hybrid feature explores a good trade-off between the detection accuracy and efficiency.Besides,we carefully design the feature to well support the streaming scenario.At last,we extend the fall detection system to be a general fast action detection system by introducing more well-designed features.In a nutshell,we fully utilize the motion information in the image and video and de-velop several novel and robust algorithms to handle and analyze the motion in the image generation and video content recoding.Quantitative experimental results demonstrate the superiority of our proposed approaches in modeling the camera motion and hierar-chically characterizing the human motion.