Investigations On Moving Objects Detection Methods Via Deep Learning-based Frameworks

Deep learning has been widely considered as one of the most popular techniques in machine learning in the past few years.The usage of deep learning for moving object detection is important in contemporary computer vision and machine learning studies.Nowadays,there are many methods proposed for solving the target objects detection problem.However,conventional methods cannot accurately handle complex situations,such as sudden illumination changes,rigid/non-rigid deformation of objects,and occlusion of target objects.Therefore,this paper proposes a novel single moving object detection method based on deep learning techniques and models updating strategies.First,deep learning models are incorporated to help to extract latent feature from video clips.Second,deep similarity learning is conducted based on ranked regression.Third,online models updating strategies are applied to enable the learned model more adaptive towards challenging circumstances in this single moving object detection task.In order to verify that the newly proposed method,a database is constructed,which contains 10 video clips and the volume reaches 1264 frames.There are totally 8 single moving object detection methods been compared.Both qualitative and quantitative analysis are applied.Dozens of comparison tests are conducted from the statistical point of view.It turns out that,the new method is more accurate and robust than other compared methods.Also,the new method demonstrates satisfied performance in handling challenging circumstances,such as sudden illumination changes,rigid / nonrigid deformation,occlusions,etc.