Study On Algorithms Of Auto Semantic Labeling For Traffic Scene

The task of image labeling is to attach each pixel to a semantic label which it belongs to.It's a very important and basic procedure of scene understanding and plays an essential role in the field of computer version.The information captured in the street scene can record and express the real traffic scene in the process of driving under driving vision.Labeling street scene video within the perspective of driving can provide important statistics and decision-making for intelligent driving equipment and urban traffic control system.In this paper,we concentrate on the traffic image and video dataset collected within driving.We built a robust model for single image labeling within multi-scale cascade structure and proposed a supervoxel-tracking method for video labeling after studying a lot of state-of-art approaches.The main work and achievements are as follows:In this paper,based on the commonly accepted observation that different semantic objects in an image at different resolutions may have different representation,we propose a novel multi-scale cascaded hierarchical model(MCHM)that substantially outperforms the state of the art image labeling methods.Our proposed approach first creates multi-resolution images from the original image to form an image pyramid and labels each image at different scale individually.Next,it constructs a cascaded hierarchical model and a feedback circle between image pyramid and labeling methods.The result from the original image labeling is used to adjust labeling parameters for images in other scale(scaled images).The labeling results from the scaled images are then fed back to enhance the original image labeling results.These naturally form a global optimization problem under scale-space condition.We further propose a desirable iterative algo rithm in order to run the model.The global convergence of the algorithm is proven through iterative approximation with latent optimization constraints.This paper proposed a supervoxel-tracking method for multi-labeling in street scene video.Supervoxel graph is a collection containing the superpixels as nodes and edges between the nodes,which can represent the temporal and spatial relationship among the superpixels.In this paper,a multi-label supervoxel graph is structured for the whole traffic scene containing different semantic label as an extension of single foreground-graph method.Each label object in the frame sequences is expressed by a collection of supervoxels.Then with the help of super-tracking method which can model the appearance of difference objects represented by related supervoxel set,the label objects in the video can be captured in long term,in which the tracking method can deal with the problem of shape deformation,color change,visual drifts,etc.Finally,in addition to programming than the multi-scale model cascade level effectiveness and supervoxel graph tracking algorithm and proof,the paper constructed in accordance with PASCAL standard contains a large amount of valid experimental driving scene dataset,and according to the corresponding demand for design and implement a range of tools for application data sets for use by other researchers to use.