Efficient Object Detection Based The Sparse Inner Product And Distributed Computing

Object detection is a key problem in multimedia processing such as imageunderstanding and video analysis. Due to the large amount of images and videos andthe limited computational resource, effective and efficient object detection ischallenging. Low computational complexity and high generalization are twoimportant goals for video object detection. Low computational complexity here meansnot only fast speed but also less energy consumption. The sliding window objectdetection method with linear support vector machines (SVMs) is a general objectdetection framework. The computational cost is herein mainly paid in complex featureextraction and inner-product based classification. Making best use of spatial-temporalcorrelation, this paper first develops an energy-saving computation method based onthe correlation of the sliding-window and a distributed object detection framework(i.e., DOD), where the process of feature extraction and classification is distributed inthe current frame and several previous frames.For classification based on inner-product, the proposed method rejects a bundleof neighboring sub-images with one inner-product operation and limited numbers ofaddition ones. The total number of multiplications and additions is equal to that of thetraditional sliding-window method. But in the proposed method, the number ofmultiplications is much smaller. With the fact that addition consumes less energy thanmultiplication, the proposed algorithm is more energy-saving and efficient.The feature extraction and classification are divided into the current frame andseveral previous frames in DOD. Only sub-feature vectors are extracted and theresponse of partial linear classifier (i.e., sub-decision value) is computed. To reducethe dimension of traditional block-based HOG (Histograms of Oriented Gradients)(BHOG) features, this work propose a cell-based HOG (i.e., CHOG) algorithm, wherethe features in one cell are not shared with overlapping blocks. Using CHOG asfeature descriptor, we develop CHOG-DOD as an instance of DOD framework.Experimental results on detection of hand, face, and pedestrian in video shows thesuperiority of the proposed method.