RGB-D Objection Detection Based On Sparse Representation

Recently, RGB-D cameras are capable of providing high quality synchronize images of both color and depth. With its advanced sensing capabilities, this technology represents an opportunity to dramatically increase the performance of object detection. However, the depth data has poor texture and far small objects always have low spatial resolution, these raise the problem of developing expressive features for the depth channel. Moreover, objects have a distinct depth contour and shape which is different from intensity domain, so finding a good way to fusion intensity image and depth image is another challenge.Up to now, most of the state-of-the-art systems derive features from intensity images, like HOG(histogram of hog), LBP etc. However, these manually designed features work in specific fields and are lack of generic. in the last couple of years, the problem of learning feature representation has been studied extensively. Most approaches have explored unsupervised learning algorithm to generate the features used for vision system, such as auto-encoders, RBMs and sparse coding. Inspiring by the HOG, we compute a histogram of sparse codes with joint dictionary learned on a set of grayscale images and depth images, the final features are the concatenation of sparse codes from multi-channels.In this way, we presents a novel approach which involves a dictionary-level fusion of intensity and depth cues, the core idea of joint dictionary learning method is to find a common sparse representation of two channels simultaneously with a learned dictionary pair.Moreover, considering the depth gradients are differs from the intensity gradients in some ways, the constrain of common representation is too strong. consequentially, we perform semi-jointed dictionary learning for pixel level features. It relaxes the strong regularization of "common sparse representation" in joint dictionary learning by calculating a linear mapping between a sparse codes pair.Our experiments on a large real-world dataset demonstrate that by utilizing the relation between depth and intensity of RGB data with learning, we achieve to yield more efficient and discriminative features.