Efficient topview person detector using point based transformation and lookup table

Nowadays due to big data revolution, image analytics is seen as a potential solution to solve different real-life problems. In this regard, one of the applications could be a person detection system. The overhead mounted camera with a wide-angle lens gives more coverage and visibility in occluded and cluttered environments than a traditional or frontal view. Person detection from the top view is a challenging task because there is variation in position, orientation, poses, body articulation and appearance of a person depending upon the position in the scene. To handle these issues an efficient method is proposed that uses different geometric transformations, concepts of perspective geometry and homography matrix as a pre-processing step. The composite transformation matrix is then used with perspective transformations to standardized the shape of the image patch containing person. At this stage, the histogram of oriented features is extracted along with our proposed five additional spatial features. These features are then fed to a linear SVM classifier for training and testing and finally, a simple effective clustering process is used to accumulate the votes of SVM to render a decision for localization of the Person in the image. To reduce the computational cost of these points based geometric and perspective transformations a Lookup table structure is used which contains pre-calculated positions of different perspective points against spatial co-ordinates. The use of lookup table and point based operations significantly reduced processing time up to 50% as compared to the previous approach which uses RHOG algorithm with image-based rotations, transformations, interpolations, and no lookup structure. The proposed method is efficient both in terms of computation and accuracy. The performance of the proposed algorithm is tested using our newly recorded dataset having more wide coverage of the scene. The performance of the developed technique shows an accuracy of 98% TDR with 3% FDR.