3D Local Spatio-temporal Ternary Patterns for Moving Object Detection in Complex Scenes

Humans possess natural cognitive vision to perceive objects in a 3D space and are able to differentiate foreground and background moving objects using their shape, colour and texture. Moving object detection is a leading and challenging task in complex scenes which involve illumination variation, blurriness, camouflage, moving background objects, etc. Inspired by human cognitive vision, a novel descriptor named 3D local spatio-temporal ternary patterns (3D-LStTP) is proposed for moving object detection. The 3D-LStTP collects multidirectional spatio-temporal information from three consecutive frames in a video by forming a 3D grid structure. The background models are constructed by using texture and colour features. The results obtained after modelling are integrated for foreground moving object detection in complex scenes. The performance of proposed algorithm is validated by conducting five experiments on Fish4Knowledge dataset, four experiments on I2R dataset and four experiments on Change Detection dataset. Qualitative and quantitative analyses are carried out on benchmark datasets. The results after investigation prove that the proposed method outperforms the state-of-the-art techniques for moving object detection in terms of ROC, TPR, FPR, Precision and F- measure.