Foreground Detection In Complex Scenarios

Foreground detection is currently one of the most active research topics in the domain of computer vision.It has a wide range of applications in human-computer interaction,intelligent security surveillance,intelligent transportation systems and etc.However,the complexity and diversity of video scenes bring many challenges,such as the recognition of dynamic background information,the background model initialization in complex scenes,the background model update with noise and foreground interference,and etc.To solve the challenges,this dissertation mainly devotes to background subspace based foreground detection algorithms in complex scenarios.The proposed algorithms model background and foreground information robustly,which helps to resist the information interference in model initialization,foreground/background classification and model update processes.More specifically,the main contributions of this dissertation are summarized as follows:1.A sparse error compensation based incremental foreground detection algorithm is proposed.The proposed algorithm estimates background subspace incrementally with improved principal component analysis(PCA)technique and fully considers the foreground sparsity,which increase the accuracy and adaptivity of our algorithm.To decrease the noise and foreground interference on the subspace update process,we design a probability sampling based sparse error compensation strategy,where the background model is updated with compensated images.The introduction of randomness increases the robustness of the proposed algorithm.2.An anti-interference error compensation based foreground detection algorithm is proposed to solve the problems of foreground positive feedback and dynamic background interference.We bring a spatial continuity constraint to foreground error estimation process and formulate the error estimation task into an objective optimization problem.With the help of non-local constraint theory,a fast optimization algorithm is proposed to efficiently solve the complex total variation optimization problem.The introduction of the spatial continuity constraint helps to separate dynamic background and foreground effectively.Then,an alpha-mating based error compensation strategy is designed,which increases the robustness of our algorithm.Finally,a background compensation template which does not rely on background subspace model is constructed.The application of the constructed template breaks the positive feedback channel of the foreground information and increases the anti-interference performance of our algorithm.3.A cross-covariance subspace based background modeling method is proposed.The cross-covariance based two dimensional principal component analysis(C2DPCA)technique is introduced into foreground detection field for the first time.Because the C2 DPCA technique could preserve all image covariance information compared with traditional 2DPCA methods,the background subspace is estimated accurately in our method.To update the background subspace in a more adaptive way,we extend the conventional batch mode C2 DPCA algorithm into an incremental one.Then,we formulate the background subspace learning and foreground estimation problems under a unified optimization framework,and design an effective optimization algorithm to solve it.The algorithm takes full advantage of the relationship between background subspace learning and foreground estimation processes,which helps to provide more accurate foreground detection results.4.An adaptive background basis selection based foreground detection method is proposed.We first design a region division based background basis matrix construction process.In the process,a multiple clustering evaluation based optimization algorithm that grades the sampled frames/blocks according to their clean degrees is proposed.The frames/blocks with high evaluation scores are adaptively selected to construct background basis matrix,which decreases the possibility of introducing foreground information into the background basis matrix.Then,a background basis matrix update process is proposed to increase the adaptivity of the background basis matrix.The proposed process selects a new background basis by balancing its clean degree and distinctiveness,which are obtained by utilizing the current background basis matrix information.When a basis fails to represent the current background information,the update is performed by replacing the ineffective basis with the newly selected one.Finally,a foreground detection framework which is compatible with the region division based background basis matrix construction and update processes is designed to detect foreground in complex scenarios.