Bayesian Models for Dynamic Scene Analysis

Abstract

In this chapter, we discuss Bayesian approaches for foreground object detection and localization in video surveillance applications. Two different sensors are used for these tasks: conventional electro-optical video cameras, and Rotating Multi-Beam (RMB) Lidar sensors. For the camera image sequences, we propose first a Markov Random Field (MRF)-based foreground extraction technique which is able to address cast shadow detection  and the exploitation of spatial coherence of the color and texture values observed in the foreground regions. Thereafter, based on the extracted foreground masks, we present a new Marked Point Process (MPP)-based method for pedestrian localization and height estimation in multi-camera systems, and give a detailed comparative evaluation of the proposed method versus a state-of-the-art technique. The last part of the chapter deals with Lidar point cloud processing where key challenges are compensating the low and inhomogeneous spatial resolution of the measurements, and various artifacts in point cloud formation caused by the rotating sensor technology. We also present here application examples including motion detection, gait-based pedestrian re-identification and activity recognition using a single RMB Lidar sensor which monitors the scene from a fixed position.