A self-organizing neural scheme for road detection in varied environments

Abstract

Detection of a drivable space is a key step in the autonomous control of a vehicle. In this paper we propose an adaptive vision based algorithm for road detection in diverse outdoor conditions. Our novel approach employs feature based classification and uses the Kohonen Self-Organizing Map (SOM) for the purpose of road detection. The robustness of the algorithm lies in the unique ability of SOM to organize information while learning diverse inputs. Features used for the training and testing of SOM are identified. The proposed method is capable of working with structured as well as unstructured roads and noisy environments that may be encountered by an intelligent vehicle. The proposed technique is extensively compared with the k-Nearest Neighbor (KNN) algorithm. Results show that SOM outperforms KNN in classification consistency and is independent to the lighting conditions while taking comparable classification time which shows that the network can also be used as an online learning architecture.

A robust road segmentation method based on graph cut with learnable neighboring link weights

“…The first column of both tables indicates the methods integrated into the graph-cut framework to estimate t-link weights and n-link weights, with the format of "t-link weights estimation method + n-link weights estimation method". Specially, CON represents the method in which n-link weights are calculated by neighboring contrast as in (4). With respect to GMMs, models are necessary to update with changes of road scenes.…”

“…Thus, it's necessary to seek other approaches to define n-link weights instead of calculating by neighboring contrast simply as in (4).…”

“…Cheng et al [3] applied mean-shift algorithm for unstructured road segmentation. Malik et al [4] used a self-organizing map to detect drivable road surfaces. Supervised methods regard road segmentation as a two-category classification problem and train classifier with labeled data to distinguish road regions.…”

A new road tracking method based on heading direction detection

“…4,5 In the bottom-up mode, the road detection can be achieved by extracting and analyzing local color, texture, edge, and other features of the image. [6][7][8][9][10][11] Various features can be selected to distinguish the road from the background using machine learning such as a support vector machine (SVM). 7 Besides, a neural network can be designed to identify roads.…”

Traversability analysis using terrain mapping and online-trained Terrain type classifier

“…In the literature, unsupervised learning feedbacks have been proved [3] to be effective implementation of on-line learning for road extraction. Also Zhou et al [4] implemented an on-line learning of the road: to determine the new training examples their algorithm looked for the largest connected road region; to establish that region, all the pixels were firstly classified in path and non-path (which turns out in a binary image); the resulting image was then processed with morphological operations to join pixels to the same region.…”