Carlo Regazzoni, Franco Oberti and Lucio Marcenaro
DIBE, University of Genoa, Genoa, Italy.
3D complex scenes can be defined as scenes where multiple non rigid and rigid objects dynamically interact by assuming relative poses within the field of view of to the observing sensor(s), Relative poses can be such that objects can be completely, partially or completely not visible in successive time instants. Object shape models for computer vision systems can be used to recognize objects at each time instant as well as to track them in time. Image features for recognition and tracking can be, in general, different. Nevertheless, ecological reasons suggest that using the same features for both tasks can be useful. In this paper, we present a shape model based on features that can be used both for tracking and recognition. The proposed shape model is based on a set of corners extracted from well-isolated objects during short temporal sequences. Well isolated objects are automatically detected by the proposed system, whose lower levels correspond to those of the system presented in [1]. The position of such objects on the ground plane can be tracked and predicted thanks to knowledge about camera calibration. Moreover, corners detected from the object silhouette of each well-isolated object contribute to build up a short-memory shape model. Each model is kept updated in time during periods when objects are isolated on the image plane; they are used as a basis for identifying and tracking non-rigid objects when occlusions occur. This can be done thanks to a fast Hough-like voting scheme that is robust to slight changes of object appearances with respect to models. The predicted ground-plane position is used to focalize the attention in image areas where the object presence has a higher probability, so allowing an average save of computational resources. Results are shown which show the validity of the approach. Image sequences proposed for the PETS workshop will be used together with other examples in indoor and outdoor conditions. Due to the objective of the workshop, in this paper we mainly explore the use of the proposed representation for non-rigid object tracking in cluttered scenes. Nevertheless, it is interesting to notice how proposed models can be used for recognition purposes in approaches similar to the ones proposed in [2]. This aspect is discussed in the paper.
References
[1] A. Tesei, A. Teschioni, C.S. Regazzoni and G. Vernazza, Long Memory Matching of interacting complex objects from real image sequences, Proc. of Conf on Time Varying Image Processing and Moving Objects Recognition 1996, pp. 283-286.
[2] D. L. Swets and J. Weng, Hierarchical Discriminant Analysis for Image Retrieval, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol 21, No. 5, May 1999, pp. 386-401.