Research On Pedestrian Detection And Depth Estimation Of Road Scene

Advanced Driver Assistance System(ADAS)is an important part of Intelligent Transport Systems,which has significant meaning to reduce traffic accidents,economic losses,people injury,and environmental pollution and improve the capacity of road transportation.As an effective method to reduce the casualties and economic losses,ADAS get more and more attention.The method of scene perception based computer vision is widely used in different kinds of Advanced Driver Assistance Systems because of its advantages of similar to human vision,low cost and generalization.In order to provide more comprehensive,accurate and timely road environment information,we proposed the solution or optimization for some key technologies in the Advanced Driver Assistance Systems based on computer vision technology in this paper.In the Advanced Driver Assistance Systems,the pedestrians detection and depth estimation of the scene is two of the most important parts,which directly related to personal safety.Therefore,this paper focuses on pedestrian detection and depth estimation of the road scene and optimizes or solves pedestrian detection and depth estimation based on computer vision technology.The main work of this paper can be summarized as follows.Firstly,some methods of pedestrian detection and depth estimation based on computer vision are introduced in this paper.Research significance and research content of pedestrian detection and depth estimation used in advanced driver assistance system are elaborated at firstly.Combined with the research hotspot and application requirement of advanced driver assistance system,we discuss the application value and significance of pedestrian detection and depth estimation based on monocular and binocular vision in advanced driver assistance system.Secondly,a pedestrian co-detection framework for binocular stereo sequences is proposed.Binocular vision and consecutive frames can provide more information than a single image.That information can be used to improve detection performance by reducing the number of candidate detection windows with low-confidence or enhancing the high-confidence candidates.To design the framework,we follow the intuition that a pedestrian has consistent appearance when observed from the same or different viewpoints.First,a baseline detector is used in both stereo images with a conservative threshold in order to expand the set of detection candidates.Before the detection process,an assisted stixel world model is computed for both left and right frames.Thus the search range of the detector is greatly reduced with the help of stixel world.Second,adjacency constrained patch matching is proposed to build correspondence between two candidates in both intra-and inter-sequence of binocular vision.Finally,we establish a mechanism to update the score of the detection aided by the corresponding candidates.Experimental results show that our framework significantly improves the performance of the evaluated baseline pedestrian approaches.Thirdly,a pedestrian detection framework aided by the fusion of information between binocular visions is proposed.In this framework,we follow the intuition that a pedestrian has consistent appearance when observed from different viewpoints.A baseline detector is used on both the left and right images with a conservative threshold in order to preserve a larger candidate set.Then adjacency constrained search based on the disparity map is applied to find the optimal matching pairs between the left and right candidate sets.After that,a mixture model of two-pedestrian detector is designed to capture the unique visual cues which are formed by two nearby pedestrians but cannot be captured by single-pedestrian detectors.Finally,an information fusion module is established to model the relationship between the single-and double-pedestrian detectors as well as to refine the final detection decisions.Compared with single image pedestrian detection,our detection framework has the potential of aggregating information from multiple images to improve the detection on individual image.At the end,a depth estimation method base on monocular vision has been proposed.The internal parameters and the external parameters need be calibrated if the binocular vision was used for depth estimation.The internal parameters and the external parameters of camera change at all times when the car which mounted with the binocular camera is running.Moreover,the ambient temperature can also affect the internal parameters and the external parameters.Therefore,the internal parameters and the external parameters cannot be calibrated precisely which make the depth estimation inaccurately.So,depth estimation with binocular vision is not practical.For these reasons,a depth estimation method based on fully convolution neural network for road scenes is proposed.Image sequences are used to train a network model for depth estimation in this method.The proposed method which avoiding the problem of binocular vision can obtain a good performance for depth estimation.