Lars Petersson

Statistical Threat Assessment for General Road Scenes Using Monte Carlo Sampling

This paper presents a threat-assessment algorithm for general road scenes. A road scene consists of a number of objects that are known, and the threat level of the scene is based on their current positions and velocities. The future driver inputs of the surrounding objects are unknown and are modeled as random variables. In order to capture realistic driver behavior, a dynamic driver model is implemented as a probabilistic prior, which computes the likelihood of a potential maneuver. A distribution of possible future scenarios can then be approximated using a Monte Carlo sampling. Based on this distribution, different threat measures can be computed, e.g., probability of collision or time to collision. Since the algorithm is based on the Monte Carlo sampling, it is computationally demanding, and several techniques are presented to increase performance without increasing computational load. The algorithm is intended both for online safety applications in a vehicle and for offline data analysis.

A new pedestrian dataset for supervised learning

This paper presents a comparative analysis of different pedestrian dataset characteristics. The main goal of the research is to determine what characteristics are desirable for improved training and validation of pedestrian detectors and classifiers. The work focuses on those aspects of the dataset which affect classification success using the most common boosting methods. Dataset characteristics such as image size, aspect ratio, geometric variance and the relative scale of positive class instances (pedestrians) within the training window form an integral part of classification success. This paper will examine the effects of varying these dataset characteristics with a view to determining the recommended attributes of a high quality and challenging dataset. While the primary focus is on characteristics of the positive training dataset, some discussion of desirable attributes for the negative dataset is important and is therefore included. This paper also serves to publish our current pedestrian dataset in various forms for non-commercial use by the scientific community. We believe the published dataset to be one of the largest, most flexible, and representative datasets available for pedestrian/person detection tasks.

Vision in and out of vehicles

At the Australian National Universitys Intelligent Vehicle Project, we are developing subsystems for: driver fatigue or inattention detection; pedestrian spotting; blind-spot checking and merging assistance to validate whether sufficient clearance exists between cars; driver feedback for lane keeping; computer-augmented vision (that is, lane boundary or vehicle highlighting on a head-up display); traffic sign detection and recognition; and human factors research aids Systems that perform such supporting tasks are generally called driver assistance systems (DAS). We believe that implementing DAS could prevent similar accidents or at least reduce their severity.

Systems integration for real-world manipulation tasks

A system developed to demonstrate integration of a number of key research areas such as localization, recognition, visual tracking, visual servoing and grasping is presented together with the underlying methodology adopted to facilitate the integration. Through sequencing of basic skills, provided by the above mentioned competencies, the system has the potential to carry out flexible grasping for fetch and carry in realistic environments. Through careful fusion of reactive and deliberative control and use of multiple sensory modalities a significant flexibility is achieved. Experimental verification of the integrated system is presented.

Large scale sign detection using HOG feature variants

In this paper we present two variant formulations of the well-known Histogram of Oriented Gradients (HOG) features and provide a comparison of these features on a large scale sign detection problem. The aim of this research is to find features capable of driving further improvements atop a preexisting detection framework used commercially to detect traffic signs on the scale of entire national road networks (1000s of kilometres of video). We assume the computationally efficient framework of a cascade of boosted weak classifiers. Rather than comparing features on the general problem of detection we compare their merits in the final stages of a cascaded detection problem where a features ability to reduce error is valued more highly than computational efficiency. Results show the benefit of the two new features on a New Zealand speed sign detection problem. We also note the importance of using non-sign training and validation instances taken from the same video data that contains the training and validation positives. This is attributed to the potential for the more powerful HOG features to overfit on specific local patterns which may be present in alternative video data.

Visibility Enhancement for Roads with Foggy or Hazy Scenes

Bad weather, particularly fog and haze, commonly obstruct drivers from observing road conditions. This could frequently lead to a considerable number of road accidents. To avoid the problem, automatic methods have been proposed to enhance visibility in bad weather. Methods that work on visible wavelengths, based on the type of their input, can be categorized into two approaches: those using polarizing filters, and those using images taken from different fog densities. Both of the approaches require that the images are multiple and taken from exactly the same point of view. While they can produce reasonably good results, their requirement makes them impractical, particularly in real time applications, such as vehicle systems. Considering their drawbacks, our goal is to develop a method that requires solely a single image taken from ordinary digital cameras, without any additional hardware. The method principally uses color and intensity information. It enhances the visibility after estimating the color of skylight and the values of airtight. The experimental results on real images show the effectiveness of the approach.

Driver assistance systems based on vision in and out of vehicles

As computer vision based systems like lane tracking, face tracking and obstacle detection mature an enhanced range of driver assistance systems are becoming feasible. This paper introduces a list of core competencies required for a driver assistance system, the issue of building in robustness is highlighted in contrast to leaving such considerations to a later product development phase. We then demonstrate how these issues may be addressed in driver assistance systems based primarily on computer vision. The underlying computer vision systems are discussed followed by an example of a driver support application for lane keeping based on force-feedback through the steering wheel.

Online stereo calibration using FPGAs

Online stereo calibration is useful in many situations where the cameras are moving relative to each other. The motion can either be intentional, as in an active stereo head, or due to vibrations, heat etc. which is commonly found in automotive applications. However, most approaches for finding the essential matrix relating the two cameras are computationally very expensive and, hence, this problem must be addressed. In this paper, we suggest deferring a large portion of the image processing onto a field programmable gate array (FPGA) since most operations can be heavily parallelized. The specific algorithm chosen to find point correspondences between the left and the right images is SIFT, which has the advantage of producing a very small number of outliers. Having few outliers is important as computing the essential matrix from point correspondences is an inherently unstable problem, particularly in the case where the cameras are nearly parallel. We present a system, which computes the computationally intensive parts of SIFT (Gaussian pyramid, Sobel etc) using an FPGA. The host computer then uses the resulting point correspondences to estimate the essential matrix with the help of a reduced model of the camera setup. On-line stereo calibration at frame rate (60Hz) is then possible without excessively loading the host computer.

Response Binning: Improved Weak Classifiers for Boosting

This paper demonstrates the value of improving the discriminating strength of weak classifiers in the context of boosting by using response binning. The reasoning is centered around, but not limited to, the well known Haar-features used by Viola and Jones (2001) in their face detection/pedestrian detection systems. It is shown that using a weak classifier based on a single threshold is sub-optimal and in the case of the Haar-feature inadequate. A more general method for features with multi-modal responses is derived that is easily used in boosting mechanisms that accepts a confidence measure, such as the RealBoost algorithm. The method is evaluated by boosting a single stage classifier and compare the performance to previous approaches

Road scene monotony detection in a fatigue management driver assistance system

Automated fatigue detection devices show much promise in combating fatigue related accidents. One aspect which hampers the introduction of these technologies is context awareness. In this paper we develop and evaluate a road scene monotony detector. The detector can be used to give context awareness to fatigue detection tools to minimise false positives. The approach could also be used by road makers to quantify monotony on fatigue prone stretches of road. The detector uses MPEG compression to measure the change in information content of the road scene over time. We show that the detector correlates highly with human identified monotonous scenes. The technique is consistent over time and applicable for day and night operation. The compression is augmented with lane tracking data to distinguish between otherwise difficult cases. The detector is integrated into a fatigue management driver assistance system.