菜鸟看论文——Free Space Compute reviews

参考：

16.Identifying Good Training Data for Self-Supervised Free Space Estimation

12.Recovering free spaceof indoor scenes from a single image

02.Real time obstacle detection in stereovision on non flat road geometry through” v-disparity

15.Ground Segmentation and Occupancy Grid Generation Using Probability Fields

In a typical traffic scenario, the following challenges and problems are faced in the analysis of the 3D environments: 1) detecting road surfaces and lane areas; 2) classification of potential obstacles above road surfaces and obtaining their 3D information; 3) classification of the roadside structures such as guardrails, lamp poles, and traffic signs.

The problem of estimating free space in structured and static environments is usually solved by exploiting properties of certain well defined structures. Two examples of free space estimation solutions are that of Hedau et al. [2] and Labayrade et al. [3]. While the first exploits the box like geometry of furniture to estimate free space in indoor scenes from a single camera image, the second uses the planar geometry of a road and identifiable lane markings to estimate the free space in urban road scenes.In unstructured or unknown environments such as forest areas, the lack of structure of the scene causes methods relying on static scene properties to fail.

The slope of the road surface shown in a road image is coupled with a lateral slope and a longitudinal slope in terms of the road direction. Compared with the longitudinal slope, the change in the lateral slope is small. Accordingly, most algorithms that estimate the slope of the road surface have typically estimated the longitudinal slope [1,7]. This paper also deals with the estimation of

LoPORS. The estimation of the lateral profile of the road surface remains a problem needing to be solved.

To account for the ever changing properties of free space in unstructured scenes, it is natural to resort to learning based systems, which usually require a training phase in which training data representing free space is used as an input to the learning algorithm.

Recent free space estimation approaches tackle this problem through self-supervision, where one classifier directly supervises input to a second classifier. The first classifier uses data it is confident about to label parts of the environment as free space; this data is then provided as input to the second classifier that extends the labeling over the whole environment. The proposed system in this paper lies within this framework, allowing long range fully autonomous free space estimation without relying on any rigid assumptions such as a planar ground or bootstrapping methods.

A variety of sensors and sensor combinations have previously been employed for free space estimation.Apart from the being expensive and power hungry, radars offer a narrow field of view and low accuracy in lateral direction [4]. LiDARS are also relatively expensive and big consumers of energy. Range measurement can alternatively be acquired using a stereo camera rig. The advantage of such system is its relatively low cost and the added bandwidth of information that a picture offers; however it suffers from drawbacks such as the requirement for calibration between the two cameras and the degradation of depth information as the depth increase.Due to their high resolution and precision, the use of vision sensors as the primary sense to environment perception became popular over the last decade. In contrast with other active sensors such as lasers or radars, vision sensors are passive one and provide the richest information. Although being sensitive to weather and illumination conditions, stereovision systems can measure the ranges to objects by a process of calculating disparities between the stereo images.

参考：

13.Complex Ground Plane Detection Based on V-disparity Map in Off-road Environment

There are three methods of ground plane detection based on 3D data: projective transformation, plane fitting and V-disparity. Projective transformation [2] [3] is just suitable for well-arranged environment. During the process of plane fitting [4] [5] the number of planes is needed, but it cannot be acquired in complex plane environment. Compared to other methods, the V-disparity algorithm is performed without using any prior knowledge of the scene appearance such as ground edges or lane-markings. Because both ground planes and obstacle can be mapped into segments in V-disparity map,it is useful for obstacle detection. It also can be used in multi-ground plane situation. Thus V-disparity algorithm can be applied in the wide range of situations in off-road environment. The existing methods based on V-disparity map have some limitations in detecting non-flat ground especially in off-road environment.

There are two kinds of applications for the V-disparity algorithm.One is to be used for detecting obstacles on the ground [7][8][9][10][11][12][13]. It is usually combination with U-disparity map to improve the result of obstacle detection [7][8][9][10].The other application is to detect ground [14][15][16][17][18][19]. U-disparity map is also used to get the detail information of ground [15].

The U-V-disparity concept was proposed by Hu [5] as an extension to Labayrade’s work [6] on the V-disparity approach. In the U-V-disparity concept, a perceived 3D environment is segmented to a 2D representation; the road is projected on V-disparity image as an oblique line, and obstacles as vertical and horizontal alignments in the V-disparity and U-disparity images (VDI and UDI) respectively, which simplify greatly the process of separation between the road surface and obstacles. The U-V disparity concept for obstacles detection and localization is robust to many factors such as illumination change, different weathers (rain, snow, fog, etc.), and road conditions (dirty areas, shadows, oil, water, etc.); this approach is used to classify the 3D road scene into relative surface planes and to characterize the road features.

Based on V-disparity theory, the enhanced V-disparity algorithm is proposed. The aim of enhanced V-disparity method is to minimize errors of ground representation which result from flat ground assumption. For the purpose of minimizing errors, sub-V-disparity map is created to represent the details of ground structure, which is useful when ground structure is complex plane. At the same time, the information of segments in sub-V-disparity is used for calculating the parameters of ground plane.

参考：

08.Online, self-supervised terrain classification via discriminatively trained submodular markov random fields

Self-Supervised Learning For Free Space Estimation: Vernaza et al. [9] also used a stereo sensor in a Markov Random Field framework to classify pixels in the image belonging belonging to the ground plane. The largest planar region is assumed to be the ground plane, and pixels belonging to it are taken as ground pixels. This training extraction method fails in scenarios where the ground plane is not the largest plane in the image. 

参考：

15.Traversability analysis for mobile robots in outdoor environments: A semi-supervised learning approach based on 3d-lidar data

06.Self-supervised monocular road detection in desert terrain

06.Probabilistic terrain analysis for high-speed desert driving

LIDAR :Sugar et al. [4]employed a 3-D LIDAR to find the occupancy probability of the environment through a semi-supervised learning approach.The robot is driven by a human operator through a safe trajectory where it collects the remission and spatial features of free space, which are used as training data for a one-class classifier. Dahlkamp et al. [5] used a 2-D LIDAR to extract training data belonging to free space using the Probabilistic Terrain Analysis (PTA) algorithm proposed in [6]. The training data is then projected to a monocular camera and used to build a color based classifier. The PTA algorithm requires unknown parameters to be learned offline using human supervision. These two systems are suitable when the properties of the robot’s operating environment resemble these of the training environment. The system presented in this paper differs from both methods in that it is totally independent of any human supervision and it does not have free parameters that need to be trained prior to deployment in a given environment.

参考：

14.Visual ground segmentation by radar supervision

Radars have also been successfully employed for selfsupervision in free space estimation. Milella et al. [7] used the echo in a radar image to identify ground patches and then projected these patches to a monocular camera coordinate frame in order to train a visual classifier. The classification was done through Mahalanobis distance thresholding.The optimal threshold is determined by constructing ROC curves on a training dataset. In their work, the radar produces training patches at a specified distance of 11.4 meters in front of the robot. Unfortunately, in some scenarios distance patches might not posses the same features as closer ones, thereby causing the latter to be classified as obstacles. The system presented in this paper gets around this problem by extracting training patches from all over the field of view of the camera.

参考：

13.A multi-baseline stereo system for scene segmentation in natural environments

09.Learning long-range vision for autonomous off-road driving

Stereo cameras are also used for self-supervised free space estimation and provide a dense 3-D representation of the scene with additional color information. 

Point cloud：Milella et al. [8] utilizes a stereo camera to extract geometric features that are used to classify voxels in a 3-D point cloud belonging to free space. The dependence on point clouds provide a coarse pixel segmentation of Ground.Furthermore,in order to create the ground model, the system needs to be initialized in an area free of obstacles. The requirement for initialization is problematic when the system fails and the human operator cannot intervene to reinitialize it. Our system does not need any special initialization and in fact can be launched inside a heavily cluttered scene.Hadsell et al.[6] also propose a stereo-based solution to segmentation in which plane fitting on point cloud data is also used. The disadvantage is that the solution requires one to specify apriori the number of planes to extract, and thus is not suitable for multi-plane environments.

参考：

07.Free space computation using stochastic occupancy grids and dynamic programming

08.Probabilistic representation of the uncertainty of stereo-vision and application to obstacle detection

Occupancy Grids：Another method to segment the ground plane is the construction of occupancy grids. The free space represented in these grids is then projected onto the image plane to segment ground pixels [7],[8]. However, these methods are sensor-specific since they require the knowledge of the sensors characteristics in order to model its uncertainty before building the occupancy grids.

参考：

15.Ground segmentation and occupancy grid generation using probability fields

Taditional line detection methods were unreliable when trying to estimate slanted lines in off-road scenes, and proposed a binary filtering algorithm that takes as an input the v-disparity image and provides as an output a filtered v-disparity image containing only slanted lines. This was coupled with a stochastic model, which uses maximum likelihood linear regression with a first order polynomial basis functions to provide point estimates of the parameters of the ground correlation line. The parameters were then used to estimate the mean of a probability distribution that describes the occupancy probability of each pixel.Then, we use the probability field generated to not only perform precise ground pixel segmentation but also to construct occupancy grids.Lastly,our method is not affected by shortcomings of traditional line detection methods on low quality v-disparity images

参考：

02.Real time obstacle detection in stereovision on non flat road geometry through” v-disparity

04.Uv-disparity: an efficient algorithm for stereovision based scene analysis

13.Complex ground plane detection based on v-disparity map in off-road environment

13.Stereo vision based traversable region detection for mobile robots using uvdisparity

V-Disparity:It transforms a disparity image to a vdisparity image by forming a 256-bin histogram of disparity values for each row of the disparity image and concatenating them vertically. For example, a 720x1280 disparity image is transformed into a 720x256 v-disparity image.Oblique and horizontal planes in the disparity image are mapped to slanted lines in the v-disparity image. Thus, detecting slanted lines in the v-disparity image is equivalent to the estimation of the ground plane in the disparity image.Compared to other methods that use stereo sensors, the v-disparity algorithm introduced in [9] does not require any initialization or knowledge about the scene. This method acts as a transformation on the image, where vertical planes are mapped to vertical lines and ground planes are mapped to slanted lines, denoted as the “ground correlation lines”. Most applications using the v-disparity algorithm tend to focus on Cartesian plane estimation instead of precise pixel segmentation. In the work of Hu et al. [4], planes are used to model Ground and obstacles in urban traffic environments.

The system provides the equation of Ground and obstacle planes in the image instead of segmenting pixels. In the work of Labayrade et al. [9], a Hough transform is used to extract lines that represent obstacles and Ground from the v-disparity image, and pixels are classified according to how well their v-d coordinates fit these lines. Since this method relies on a Hough transform for line detection, it could only detect the ground plane correctly provided that the ground correlation line is the most dominant line in the v-disparity image. This method was improved by Yiruo et al [10], where a moving window is used over the v-disparity image, and the segmented lines that are detected in that window are used to find the ground plane. The results show an increase in performance over the method of [9], but it is focused on traversable region segmentation, which has the inconvenience of leaving large areas around obstacles unidentified. In the work of Xiaozhou et al. [11], Labayrade’s work [9] is further improved by utilizing the u-disparity image in order to remove obstacles from the disparity map before constructing the v-disparity map and segmenting the ground plane. This method provides acceptable results in urban traffic environment but tends to fail if un-textured patches of pixels are located on the ground plane.

参考

15.Identifying Good Training Data for Self-Supervised Free Space Estimation

This paper proposes a novel technique to extract training data from free space in a scene using a stereo camera.The proposed technique exploits the projection of planes in the v-disparity image paired with Bayesian linear regression to reliably identify training image pixels belonging to free space in a scene. Unlike other methods in the literature,the algorithm does not require any prior training, has only one free parameter, and is shown to provide consistent results over a variety of terrains without the need for any manual tuning. The proposed method is compared to two other data extraction methods from the literature. Results of Support Vector classifiers using training data extracted by the proposed technique are superior in terms of quality and consistency of free space estimation. Furthermore, the computation time required by the proposed technique is shown to be smaller and more consistent than that of other training data extraction methods.The novel training data extraction algorithm presented in this paper utilizes the properties of the projection of the ground on the v-disparity image, and is able to extract training pixels even if the ground is not planar.