Ion Giosan

UV disparity based obstacle detection and pedestrian classification in urban traffic scenarios

High accuracy pedestrian detection plays an important role in all intelligent vehicles. This paper describes a system for detecting the obstacles in front of the vehicle and classifying them in pedestrians and non-pedestrians. It acquires the traffic scenes using a low-cost pair of gray intensities stereo cameras. A SORT-SGM stereo-reconstruction technique is used in order to obtain high density and accuracy in stereo-reconstructed points. First, the road plane is computed using the V disparity map and then the obstacles are determined by analyzing the U disparity map. Size related and histogram of oriented gradient based on gray levels features are used for describing each pedestrian hypothesis. A principle component analysis on the features is used for their selection and projection in a relevant space. Different SVM classifiers are trained considering the relevant features on large pedestrian and non-pedestrian image sets. A comparison between them is finally performed for selecting the one that achieves the best classification score.

Real time stereo vision based pedestrian detection using full body contours

There are many approaches to pedestrian detection in collision avoidance systems depending on the sensors (visible light, thermal infrared, RADAR, LASER scanner) used for acquiring the data and the features (depth, shape, motion) used for detection. In this paper we present a method for shape based pedestrian detection in traffic scenes using a stereo vision system for acquiring the image frames and a contour matching technique for classifying the scene objects as belonging or not to the pedestrian class. The 3D information is used for determining the foreground points of each 2D object and a contour extraction and merging algorithm is performed on these points. A 2D image filtering is performed and edges are extracted and used for objects contours refinement. A hierarchy of pedestrian full body contours and a matching technique are used for classifying the extracted objects contours from the scene.

Superpixel-based Obstacle Segmentation from Dense Stereo Urban Traffic Scenarios Using Intensity, Depth and Optical Flow Information

Obstacle detection is a necessary task in every driving assistance system. An accurate obstacle segmentation is very important for further processing tasks that are using the obstacle ROI as input, like obstacle classification. This paper presents a real time approach for obstacle segmentation from traffic scenarios, based on superpixels clustering. A pair of gray levels stereo-cameras is used for scene image acquisition. The stereo-reconstruction uses a sub-pixel level optimized semi-global matching (SORT-SGM) resulting in a very accurate 3D points map. Optical flow is computed using a Lukas-Kanade pyramidal approach. A novel paradigm integrating intensity, depth and optical flow information on superpixels is used for obstacle segmentation. SLIC superpixels are computed first based on intensity information. Multiple features are computed for each superpixel and used for clustering superpixels in obstacles. Depth cues are used for clustering the superpixels in obstacles and then optical flow information refines the obstacles clusters based on their motion. A qualitative and quantitative evaluation of the proposed approach and a comparison with other obstacle detection technique are finally presented.

Pedestrian detection in the context of multiple-sensor data alignment for far-infrared and stereo vision sensors

Multiple sensor systems are extremely used in autonomous driving for providing increased object detection accuracy. We present a multiple sensor based pedestrian detection system that combines aggregated channel features classifiers trained on images captured with two types of sensors: far infrared and stereovision sensors. We developed a spatio-temporal data alignment between the two sensorial systems. For the temporal alignment we used an original camera response timing model for free running cameras in order to align the infrared image with grayscale intensity images captured by trigger-based cameras. The spatial data alignment is done by performing the extrinsic parameters calibration of the infrared camera in the ego car world coordinate system which is also the reference for the stereovision sensors. Based on the aligned sensorial data we developed a classification fusion mechanism for combining infrared and grayscale detections on a unified pedestrian detection stream. We obtain an increased accuracy showing that the two detectors complete each other.

Urban traffic dense-stereo obstacle classification using boosting over visual codebook features

Every driving assistance system should have an obstacle classification module. Its main role is to accurately classify obstacles within a set of predefined classes. This paper presents a real-time dense-stereo based obstacle classification system that integrates visual codebook features like HOG, LBP and texton descriptor types in a powerful classifier. The system classifies the obstacles in four main classes: cars, pedestrians, poles/trees and other obstacles. The system acquires the image scenes using a pair of gray level stereo video-cameras. A combined approach using both 2D intensity and 3D depth information is firstly used for accurate obstacle segmentation. Then, the visual codebook features are extracted for a large set of obstacles with manually labeled classes and used for training a robust boosting classifier. The comparative classification results with an approach based on a random forest classifier trained on a relevant feature set show a considerable improvement, especially for the class of other obstacles.

Animal detection from traffic scenarios based on monocular color vision

This paper presents a system capable of detecting various large sized wild animals from traffic scenes. Visual data is obtained from a camera with monocular color vison. The goal is to analyze the traffic scene image, to locate the regions of interest and to correctly classify them for finding the animals that are on the road and might cause an accident. A saliency map is generated from the traffic scene image, based on intensity, color and orientation features. The salient regions of this map are considered to be regions of interest. A database is compiled from a large number of images containing different four-legged wild animals. Relevant features are extracted from these and are used to train Support Vector Machine classifiers. These classifiers provide an accuracy of above 90% and is used to predict whether or not the selected regions of interest contain animals. If one of the regions is classified as containing an animal, a warning can be signaled.

Real-time object detection using a sparse 4-layer LIDAR

The robust detection of obstacles, on a given road path by vehicles equipped with range measurement devices represents a requirement for many research fields including autonomous driving and advanced driving assistance systems. One particular sensor system used for measurement tasks, due to its known accuracy, is the LIDAR (Light Detection and Ranging). The commercial price and computational intensiveness of such systems generally increase with the number of scanning layers. For this reason, in this paper, a novel six step based obstacle detection approach using a 4-layer LIDAR is presented. In the proposed pipeline we tackle the problem of data correction and temporal point cloud fusion and we present an original method for detecting obstacles using a combination between a polar histogram and an elevation grid. The results have been validated by using objects provided from other range measurement sensors.

Superpixels in Pedestrian Detection from Stereo Images in Urban Traffic Scenarios

Pedestrian detection is a common task in every driving assistance system. The main goal resides in obtaining a high accuracy detection in a reasonable amount of processing time. This paper proposes a novel method for superpixel-based pedestrian hypotheses generation and their validation through feature classification. We analyze the possibility of using superpixels in pedestrian detection by investigating both the execution time and the accuracy of the results. Urban traffic images are acquired by a stereo-cameras system. A multi-feature superpixels-based method is used for obstacles segmentation and pedestrian hypotheses selection. Histogram of Oriented Gradients features are extracted both on the raw 2D intensity image and also on the superpixels mean intensity image for each hypothesis. Principal Component Analysis is also employed for selecting the relevant features. Support Vector Machine and AdaBoost classifiers are trained on: initial features and selected features extracted from both raw 2D intensity image and mean superpixels intensity image. The comparative results show that superpixelsbased pedestrian detection clearly reduce the execution time while the quality of the results is just slightly decreased.

Shape improvement of traffic pedestrian hypotheses by means of stereo-vision and superpixels

Shape is a powerful descriptor frequently used in pedestrian detection process. This paper presents a novel stereo and superpixel-based approach for extracting high quality shapes of pedestrian hypotheses from urban traffic scenarios. Gray-levels stereo-vision images of traffic scenes are acquired, high quality stereo-reconstruction and optical flow algorithms are used for computing the depth and motion information. Superpixels are extracted using the intensity images and clustered in different obstacles by a novel paradigm that fuses intensity, depth and motion information. Pedestrian hypotheses are defined as a subset of the scene obstacles obtained by imposing human-specific geometric constraints. A contour tracing algorithm is used for extracting a continuous contour that defines the shape of each pedestrian hypothesis. A comparison between the contours quality of pedestrian hypotheses obtained by this stereo and superpixel approach and another approach based only on stereo-reconstructed points grouping shows improvements in both object shape description and area coverage. Improvements in shape description will increase the accuracy of any further pedestrian detection processes that use pattern matching techniques.

Modeling the behavior of large scale reasoning systems using clustering and regression

Modeling the performance of large scale systems is the core idea of this paper.We focus on modeling the performance specific behavior of LarKC 1- The Large Knowledge Collider a platform for large scale integrated reasoning and Web-search. A set of instrumentation and monitoring tools are employed to collect metrics related to execution time, resources, and specific platform measurements like running workflows and plug-ins. Our method performs machine learning on top of instrumented data and tries to find relations between input defined metrics and output metrics that describe the instrumentation observations of the LarKC platform, plug-ins or workflows. The proposed method is a combination of clustering and regression techniques.