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.

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 (1000's 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 feature's 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.

This paper demonstrates a method of increasing the quality of weak classifiers in the boosting context by using improved response modelling. The new method improves upon the results of a recent response binning approach proposed by Rasolzadeh et al. [1]. For experimental purposes the improved method is applied to the familiar Haar features as used by Viola and Jones in their face/pedestrian detection systems. However, the methods benefits are general and therefore not restricted to this particular feature type. Unlike many previous methods, this method is suitable for modelling multi-modal responses and is highly resistant to overfitting. It does this by adaptively choosing suitable support regions around the values taken by the standard response binning method. More accurate models are produced, with particular improvement around the final decision boundary. It is shown that the new method can be trained with one tenth of the training data required to achieve similar results on previous methods. This substantially lowers the overall training time of the system. The method's ability to consistently produce better hypotheses over a variety of pedestrian detection tasks is shown.

Abstract-This paper presents a fast Histogram of Oriented Gradients (HOG) based weak classifier that is extremely fast to compute and highly discriminative. This feature set has been developed in an effort to balance the required processing and memory bandwidth so as to eliminate bottlenecks during run time evaluation. The feature set is the next generation in a series of features based on a novel precomputed image for HOG based features. It contains features which are more balanced in terms of processing and memory requirements than its predecessors, has a larger and richer feature space, and is more discriminant on a per feature basis.In terms of computational complexity it is a heterogeneous feature set. I.e. it has fast and slow variants. In order to optimize our feature selections between the faster and slower features available we implement a recently proposed modification to the RealBoost feature selection rule. This modification provides an additional means to balance processing and memory bandwidth on ordinary PC architectures. This feature set is suitable for use within typical boosting frameworks. It is compared to Haar and Rectangular HOG features, as well the related feature HistFeat. The new feature set contains two variants, LiteHOG and LiteHOG+, which we compare. Both LiteHOG and LiteHOG+ features show promising results on road sign and pedestrian detection tasks.

No abstract

This paper presents various opthisations that can be applied to the Sum of Absolute Differences (SAD) correlation algorithm for automated landmark detection. This has applications in mobile robotic navigation and mapping. We show how some assumptions about the environment and the generic form of strong landmarks selected by the SAD correlation algorithm have led to the development of an algorithm to enable near real time selection of strong landmarks from visual information.The landmarks that have been selected from a series of frames using our optimisations are shown to be stable through the image sequence, demonstration the scale invariance of the landmarks that are selected by the SAD correlation algorithm. 0-78034232-3/04/$17.00 02004 IEEE

Many state-of-the-art general object detection methods make use of shared full-image convolutional features (as in Faster R-CNN). This achieves a reasonable test-phase computation time while enjoys the discriminative power provided by large Convolutional Neural Network (CNN) models. Such designs excel on benchmarks 1 which contain natural images but which have very unnatural distributions, i.e. they have an unnaturally high-frequency of the target classes and a bias towards a "friendly" or "dominant" object scale. In this paper we present further study of the use and adaptation of the Faster R-CNN object detection method for datasets presenting natural scale distribution and unbiased real-world object frequency. In particular, we show that better alignment of the detector scale sensitivity to the extant distribution improves vehicle detection performance. We do this by modifying both the selection of Region Proposals, and through using more scaleappropriate full-image convolution features within the CNN model. By selecting better scales in the region proposal input and by combining feature maps through careful design of the convolutional neural network, we improve performance on smaller objects. We significantly increase detection AP for the KITTI dataset car class from 76.3% on our baseline Faster R-CNN detector to 83.6% in our improved detector.

Recent work [1], has shown that improving model learning for weak classifiers can yield significant gains in the overall accuracy of a boosted classifier. However, most published classifier boosting research relies only on rudimentary learning techniques for weak classifiers. So while it is known that improving the model learning can greatly improve the accuracy of the resulting strong classifier, it remains to be shown how much can yet be gained by further improving the model learning at the weak classifier level. This paper derives a very accurate model learning method for weak classifiers based on the popular Haar-like features and presents an investigation of its usefulness compared to the standard and recent approaches. The accuracy of the new method is shown by demonstrating the new models ability to predict ROC performance on validation data. A discussion of the problems in learning accurate weak hypotheses is given, along with example solutions. It is also shown that a previous simpler method can be further improved. Lastly, we show that improving model accuracy does not continue to yield improved overall classification beyond a certain point. At this point the learning technique, in this case RealBoost, is unable to make gains from the improved model data.The method has been tested on pedestrian detection tasks using classifiers boosted using the RealBoost boosting algorithm. A subset of our most interesting results is shown to demonstrate the value of method.