Angle Spectrum for Estimation of Trajectory Deviation Using Combined Tracking and Neural Network Labeling Nikolaos Doulamis Vassilios Vescoukis Andeas Georgopoulos National Technical University of Athens 9, Heroon Polytechniou, 157,73 Zografou, Athens Greece Tel: +30210772-2678 ndoulam@cs.ntua.gr National Technical University of Athens 9, Heroon Polytechniou, 15773, Zografou, Athens Greece +30210-772-1688 v.vescoukis@cs.ntua.gr National Technical University of Athens 9, Heroon Polytechniou, 15773, Zografou, Athens Greece +30210-772-2675 drag@central.ntua.gr ABSTRACT In this paper, we combine rigid motion-based tracking algorithms and non-linear identification methods for automatic detecting and tracking vehicles’ trajectory in roadways. In addition, we introduce the concept of the angle spectrum for determining the deviation of a vehicle trajectory from the ideal trace, provided by surveyor engineers. Motion–based tracking is implemented through frame differencing and advanced non-linear convolution filters such as the morphological opening by reconstruction. However, motion based tracking suffers from noise, occlusions and the fact that the detected moving region may contains more than one foreground objects (e.g., a vehicle approach another vehicle). For this reason, a neural network-based classification scheme is adopted in this paper for identifying foreground/background objects. The neural network models the colour and texture properties of the detected moving objects. Fusion algorithm are then exploited which it combine the output of the neural network classifier and the output of the motion- based tracking for efficiently detecting the vehicles trajectory. In the following, we introduce the concept of the angle spectrum which estimates the deviation between two curves, i.e., the vehicle trajectory and the ideal trace. The angle spectrum is computed through quantization of the polar coordinate space, adopted for the curve representation along with novel matching schemes. Experimental results are presented, which indicate the performance of the proposed method in real file environments. Categories and Subject Descriptors I.4.8 [Image Processing and Computer Vision: Scene Analysis]-tracking; I.2.10 [Artificial Intelligence: Vision and Scene Understanding]—motion, video analysis. General Terms Algorithms, Design, Theory. Keywords Motion-based tracking, automatic neural-based labeling, curve spectrum. 1. INTRODUCTION Event analysis in video sequences is a critical aspect for many applications. Examples include, but not limited to, i) detection of actions in sport sequences, such as soccer matches [1], ii) cognitive video supervision [2], such as monitoring of human activities [3] iii) methods and tools for indentifying and tracking “targets” and scenes of interest in broadcasting applications to allow for a personalized video streaming [4] and iv) traffic surveillance systems [5]. In addition, the recent explosion of the amount of video data, being captured, shared, stored and transmitted, poses great needs in semantic indexing and querying of the video databases [6]. Methods and tools for event detection in video sequences are very important for semantic tagging of video streams. Towards this direction, Large-Scale Concept Ontology for Multimedia (LSCOM) has defined a set of concepts for activities, covering a broad range of video events such as airplane flying, car crash, riot, people marching, and so on [7]. One interesting application for event detection in videos is to monitor vehicles’ trajectory in roadways and estimate the deviation of such trajectory from the designed one. Currently, civil and survey engineers manually observe vehicle trajectories and compare them with the design trajectory in order to improve road conditions and therefore traffic safety. However, such manual process is very tedious, time consuming and error prone and its limitations can be overcome through the use of computer vision and video processing techniques. Furthermore, vehicle identification and detection is also very useful towards semantic content characterization. For most traffic surveillance systems, three major stages are used to estimate desired traffic parameters, i.e., vehicle detection, tracking, and classification. For vehicle detection, most methods assume that the camera is static and then desired vehicles can be detected by image differencing [8], [9]. In particular, in [9], entropy is used as an underlying measurement to calculate traffic flows and vehicle speeds, while in [8] a region-based approach is adopted to track and classify vehicles based on the establishment of correspondences between regions and vehicles. However, these methods are very sensitive to noise. For this reason, a self adaptive background extraction method is proposed in [10] for vehicle tracking. In this approach, vehicles are located by projecting the difference image onto the edge map. The main drawback of the above mentioned methods is that they cannot work accurate in case of vehicle occlusion. This is due to the fact that if a vehicle is hidden either from another vehicle or the background it is very difficult for a motion-based tracking Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. AREA’08, October 31, 2008, Vancouver, British Columbia, Canada. Copyright 2008 ACM 978-1-60558-318-1/08/10...$5.00. 25