Research On Moving Object Detecton And Tracking Techniques

Moving object detection and tracking is an important research domain in computer vision. Related techniques have been broadly applied in battlefield surveillance, video monitoring, image compression, image retrieve, human-computer interaction and so on. Many research works have been done in last several decades, moving object detection and tracking has achieved great progress. However, moving object detection and tracking system in general sense is still incomplete, more robust core algorithms are necessary if one wants to obtain more steady and practical systems. Nowadays, all kinds of moving object detection and tracking methods emerge endlessly. This thesis mainly focuses on part of moving object detection and tracking methods for further study. Main works of this thesis include:(1) In order to depress the negative effect of dynamic background and illumination changing to moving object detection in complex scene, the theory of sequential kernel density approximation is introduced to background modeling. We design a background subtraction algorithm based on this model. Then we boost the robustness of the moving object detection by combining pixel-level, region-level and frame-level discrimination rules.(2) Focusing on the disturbance of moving cast shadow, a Bagging-ensemble-based moving cast shadow removal method is proposed. We collect shadow discrimination features from multiple shadow discrimination models. A shadow detector is trained by employing boosting-pruning-Bagging-ensemble-based learning framework. The shadow detector can select effective shadow discrimination features and be updated online adaptively.(3) Considering the requirement of moving object tracking algorithm should adapt to the changing scene, a revised color-spatial mixture of Gaussian (SMOG) is proposed. We recognize and remove the introduced background modes in tracking initialization by calculating the spatial-color joint distance between each Gaussian mode and the object local background. We also consider the joint distance as a confidence of the discriminative power of each mode, and introduce the confidence into the similarity measure function, dynamically update these confidences in the tracking process to adapt changing background.(4) Aiming to reduce the distraction of object shape variance, partial occlusion and clutter background, a novel multiple-features-fusion based object tracking algorithm is given. We described the color, texture, edge and motion feature of the object with a united histogram model to alleviate the affection of object deformation and partial occlusion, we fused these features in the framework of auxiliary particle filter to conquer the distractions in the complex background.(5) Considering the requirement of high precision and efficiency in active vehicle safety system, an eye state recognition based driving fatigue detection method is given. We track human face by combining both adaboost face detector and mean-shift tracker to improve the robustness of face tracking. A gray morphological filter based eye detection method is given. Then many simple but effective image features are employed to recognize the eye state. Finally, fatigue detection is realized by judging the consecutive eye state.