John-Paul Renno

An object-based comparative methodology for motion detection based on the F-Measure

The majority of visual surveillance algorithms rely on effective and accurate motion detection. However, most evaluation techniques described in literature do not address the complexity and range of the issues which underpin the design of a good evaluation methodology. In this paper, we explore the problems associated with both the optimising the operating point of any motion detection algorithms and the objective performance comparison of competing algorithms. In particular, we develop an object-based approach based on the F-Measure-a single-valued ROC-like measure which enables a straight-forward mechanism for both optimising and comparing motion detection algorithms. Despite the advantages over pixel-based ROC approaches, a number of important issues associated with parameterising the evaluation algorithm need to be addressed. The approach is illustrated by a comparison of three motion detection algorithms including the well-known Stauffer and Grimson algorithm, based on results obtained on two datasets.

Adaptive eigen-backgrounds for object detection

Most tracking algorithms detect moving objects by comparing incoming images against a reference frame. Crucially, this reference image must adapt continuously to the current lighting conditions if objects are to be accurately differentiated. In this work, a novel appearance model method is presented based on the eigen-background approach. The image can be efficiently represented by a set of appearance models with few significant dimensions. Rather than accumulating the necessarily enormous training set to generate the eigen model, the described technique builds and adapts the eigen-model online evolving both the parameters and number of significant dimension. For each incoming image, a reference frame may be efficiently hypothesized from a subsample of the incoming pixels. A comparative evaluation that measures segmentation accuracy using large amounts of manually derived ground truth is presented.

Occlusion analysis: Learning and utilising depth maps in object tracking

Complex scenes such as underground stations and malls are composed of static occlusion structures such as walls, entrances, columns, turnstiles and barriers. Unless this occlusion landscape is made explicit such structures can defeat the process of tracking individuals through the scene. This paper describes a method of generating the probability density functions for the depth of the scene at each pixel from a training set of detected blobs, i.e., observations of detected moving people. As the results are necessarily noisy, a regularization process is employed to recover the most self-consistent scene depth structure. An occlusion reasoning framework is proposed to enable object tracking methodologies to make effective use of the recovered depth.

Application and Evaluation of Colour Constancy in Visual Surveillance

The problem of colour constancy in the context of visual surveillance applications is addressed in this paper. We seek to reduce the variability of the surface colours inherent in the video of most indoor and outdoor surveillance scenarios to improve the robustness and reliability of applications which depend on reliable colour descriptions e.g. content retrieval. Two well-known colour constancy algorithms - the Grey-world and Gamut-mapping - are applied to frame sequences containing significant variations in the colour temperature of the illuminant. We also consider the problem of automatically selecting a reference image, representative of the scene under the canonical illuminant. A quantitative evaluation of the performance of the colour constancy algorithms is undertaken

Evaluation of MPEG7 color descriptors for visual surveillance retrieval

This paper presents the results to evaluate the effectiveness of MPEG 7 color descriptors in visual surveillance retrieval problems. A set of image sequences of pedestrians entering and leaving a room, viewed by two cameras, is used to create a test set. The problem posed is the correct identification of other sequences showing the same person as contained in an example image. Color descriptors from the MPEG7 standard are used, including dominant color, color layout, color structure and scalable color experiments are presented that compare the performance of these, and also compare automatic and manual techniques to examine the sensitivity of the retrieval rate on segmentation accuracy. In addition, results are presented on innovative methods to combine the output from different descriptors, and also different components of the observed people. The evaluation measure used is the ANMRR, a standard in content-based retrieval experiments.

Performance evaluation in visual surveillance using the F-measure

Majority of visual surveillance algorithms rely on effective and accurate motion detection.However,most evaluation techniques described in literature do not address the complexity and range of the issues which underpin the design of good evaluation methodology.In this paper we explore the problems associated with optimising the operating point of motion detection algorithms and objective performance evaluation. We emphasise advantages of object-based approach, examine problems associated with ROC optimisation and propose a solution based on F-measure optimisation.Finally, we describe a motivated and comprehensive methodology for evaluation of motion detection within the wider context of a surveillance system.

Object Classification in Visual Surveillance Using Adaboost

In this paper, we present a method of object classification within the context of visual surveillance. Our goal is the classification of tracked objects into one of the two classes: people and cars. Using training data comprised of trajectories tracked from our car-park, a weighted ensemble of Adaboost classifiers is developed. Each ensemble is representative of a particular feature, evaluated and normalised by its significance. Classification is performed using the sub-optimal hyper-plane derived by selection of the N-best performing feature ensembles. The resulting performance is compared to a similar Adaboost classifier, trained using a single ensemble over all dimensions.

Shadow Classification and Evaluation for Soccer Player Detection

In a football stadium environment with multiple overhead floodlights, many protruding shadows can be observed originating from each of the targets. To successfully track individual targets, it is essential to achieve an accurate representation of the foreground. Unfortunately, many of the existing techniques are sensitive to shadows, falsely classifying them as foreground. In this work an unsupervised learning procedure that determines the RGB colour distributions of the foreground and shadow classes of feature data is proposed. A novel skelatonisation and spatial filtering process is developed for identifying components in the foreground segmentation that are most-likely to belong to each class of feature. A pixel classification mechanism is obtained at by approximating both classes of feature data by N Gaussian parametric models. To assess our technique’s performance and reliability, a comparison is made with other published works.

Towards plug-and-play visual surveillance: learning tracking models

In typical visual surveillance implementations, observations of scene objects are extracted as regions of moving pixels identified by pixel differencing based motion detection algorithms. These observations are tracked to establish their temporal coherence by updating a state vector describing the projected 2D width and height as well as image trajectory. Such an approach is particularly vulnerable to fragmentation and occlusion process as there is essentially no appearance model. The objective of this work is to develop simple but highly discriminatory models of scene objects which indirectly use the depth of the object to model its projected width and height. Rather than relying on a time-consuming, labour-intensive and expert-dependent calibration procedure to recover the full image to ground-plane homography, the system relies on a simple learning procedure involving watching several hundred objects entering, passing through and leaving the monitored view volume to recover the relationship between the projected 2D width and height of an object and its image position and visual motion.