RGBD Saliency Detection And Application In Pedestrian Detection

Human visual system can efficiently process the massive visual information and detect the important visual regions,in academic words,the salient subsets in a scene.Saliency detection is expected to electronically duplicate the abilities of human vision system to process big visual data.The detected regions with their level of saliency can be used by other image analysis tasks.Thanks to the development of depth capture and display technologies,there are various emerging applicants for RGBD content.This thesis focuses on RGBD saliency detection and its application in pedestrian detection.The main work are as follows:(1)Depth images produced from existing 3D sensors are often with noises.In this thesis,we propose a saliency detection scheme balancing depth image details based on convolutional neural network.Specifically,we first generate coarse depth-induced saliency cues which are careless of depth details.Then,fine details from RGB image are progressively embedded into the feature maps to refine the final prediction.In this way,we take both bottom-up and top-down cues together with spatial layout into account and achieve better saliency modeling results.Experiments on five public datasets show the superiority of the proposed method.(2)Considering the compensation relationship between depth data and RGB data,this thesis proposes to explicitly model the compensation relationship with the residual compensation module.By utilizing skip-connection,the input depth data are guided to learn complementary features.To catch different level interactions between RGB and depth,we apply a multi-level architecture to learn multi-level compensation relationships.The final prediction part combines different level features.Experiments show that this method can deal with RGBD saliency detection well with high precision.(3)In order to apply saliency detection in other tasks,this thesis proposes a pedestrian detection method based on features extracted by saliency model.The RGBD saliency model is used to extract and fuse the features of color image and depth image,the pedestrian detection task is then achieved by a simple additional detection module.The features extracted from saliency detection model contain abundant visual cues.Experiments based on open-access dataset demonstrate that the saliency detection features can be used by pedestrian detection tasks.In summary,this thesis contains saliency detection and its application in pedestrian detection task on RGBD image.In order to cope with the limitation of depth image and the fusion problem of multi-source information,the corresponding solutions are proposed.We also demonstrate the value of salient features in pedestrian detection task.