In this paper we demonstrate how the interaction between innovative methods in the field of computer vision and methods for multi-spectral image classification can help in extracting detailed land-cover / land-use information from Very High Resolution (VHR) satellite imagery. We introduce the novel concept of "geometric activity images", which we define as images encoding the strength of the relationship between a pixel and surrounding features detected through dedicated computer vision methods. These geometric activity images are used as alternatives to more traditional texture images that better describe the geometry of man-made structures and that can be included as additional information in a non-parametric supervised classification framework. We present a number of findings resulting from the integration of geometric activity images and multi-spectral bands in an artificial neural network classification. The geometric activity images we use result from the use of a ridge detector for straight line detection, calculated for different window sizes and for all multi-spectral bands and band-ratio images in a VHR scene. A selection of the most relevant bands to use for classification is carried out using band selection based on a genetic algorithm. Sensitivity analysis is used to assess the importance of each input variable. An application of the proposed methods to part of a Quickbird image taken over the suburban fringe of the city of Ghent (Belgium) shows that we are able to identify roads with much higher accuracy than when using more traditional multi-spectral image classification techniques.
In this paper, we examine sensor specific distributions of local image operators (edge and line detectors), which describe the appearance of people in video sequences. The distributions are used to describe a probabilistic articulated motion model to track the gestures of a person in terms of arms and body movement. The distributions are based on work of Sidenbladh where general distributions are examined, collected over images found on the internet. In our work, we focus on the statistics of one sensor, in our case a standard webcam, and examine the influence of image noise and scale. We show that although the general shape of the distributions published by Sidenbladh are found, important anomalies occur which are due to image noise and reduced resolution. Taking into account the effects of noise and blurring on the scale space response of edge and line detectors improves the overall performance of the model. The original distributions introduced a bias towards small sharp boundaries over large blurred boundaries. In the case of arms and legs which often appear blurred in the image, this bias is unwanted. Incorporating our modifications in the distributions removes the bias and makes the tracking more robust.