Abstract—Visual surveillance is an active area of research topic. Data that is collected from these cameras have to be monitored by human operators manually for long durations which is not feasible in real time and may lead to inaccurate results. Recorded videos are analyzed only when any unwanted event occurs that may help for recovery and not avoidance. Intelligent video surveillance requires algorithms that are fast, robust and reliable during various phases such as object & shadow detection, classification, tracking, and event analysis. This paper presents metadata based object & shadow detection, classification system for video analytics. Meta data of key frames is stored in the form of database, that helps for object & shadow detection. Gaussian white noise is used for background modeling. Convex non overlapped blobs are identified using LoG (Laplacian of Gaussians). Four channels in four color spaces are used for better removal of shadows. Shadow boundary is detected by growing a user specified shadow outline on an illumination-sensitive image among the four channels. For Object detection we used canny edge detection and Harris corner point detection for detecting the objects, Bag of visual words with SIFT descriptors for extraction of features and KNN classifier for classification. The proposed method is tested on sample CCTV camera video and results are analyzed using performance measures.

Index Terms—key frames, meta data, shadow removal, color space, bag-of-visual-words, KNN classifier

Cite: S. Vasavi and V. Srinivasa Rao, "Metadata Based Object Detection and Classification Using Key Frame Extraction Method," Journal of Image and Graphics, Vol. 3, No. 2, pp. 90-95, December 2015. doi: 10.18178/joig.3.2.90-95