FCM Based Enhanced Approach for Object Extraction from Videos

Abstract

In Video segmentation huge number of research progress has done, which includes variety of algorithms for specific applications. Videos cover wide variations in object with multiple frames capturing different pose, illumination and other variations. This diverse information can be aggregated together for efficient object segmentation. Till now, it remain challenging problem to accurately extract the target object from the video because of the variations of intensity, view, brightness, motion and other complex backgrounds. Also, accuracy and fast object tracing and object’s motion updating is necessary. In this work, we proposed a method to extract object from video sequence. It involves tracking through training phase, test phase then update the appearance model. To do this, we start with an initialized location of the object for the first frame. We do simple tracking for the first 4 frames, and use the corresponding target-background regions to train the discriminative appearance model. For each frame, we first use the appearance model to get a target confidence map, and then find the target candidate with the largest confidence. The target-around region is saved for updating appearance model every several frames. On the computed confidence map saved images, simple threshold segmentation method is applied and segmented objects are saved. Then on segmented object, apply Fuzzy C-Means clustering. This will result in improved segmentation quality. Output is shown as the segmented images of object in video sequence.

Keywords—Segmentation; FCM; PSNR; MSE; Mean-Shift; SLIC