No abstract

This paper explores a discriminative part-based approach for recognising people in video. It uses many regions to model the background and foreground and a random forest for classification. The objective is to overcome the limitations of more holistic approaches that try to recognise people as a single region with the consequential need to segment each person as one representation. Attributes of each blob, their relationships and variation over video frames are argued to be useful features for discrimination. In this paper the attributes of each blob are considered as a first step in the recognition process. We evaluate our approach through a comparison of three state of the art classifiers: Bagging, Adaboost and a Multilayer Perceptron (MLP), with the Random Forest (RF) using 10 fold cross validation. A detailed statistical analysis shows that the random forest classifier is more accurate compared to the other methods in terms of discrimination between regions describing people and those of the background.

The key issue addressed by this paper is the necessity to devise performance evaluation measures for systems that integrate multiple cues for tracking in video sequences. We propose a generic evaluation approach that can be implemented in systems that perform higher-level people tracking by integrating multiple low-level features extracted from the video data. Two new measures: video sequence accuracy (VSA) and voting average measure (VAM), are introduced and explained by using the two fundamental image processing techniques of edge and optical flow detection. The effectiveness of the approach is demonstrated using a set of real video sequences with ground truth.