Weakly Supervised Domain Adaptation for Video Surveillance in Large-scale Camera Networks

With the growing installation of surveillance video cameras in both private and public areas, closed-circuit TV has evolved from single-camera to multiple-camera systems, and more re- cently to large-scale camera networks. Current surveillance applications such as those em- ployed in shopping malls and residential buildings consist of multiple-camera networks of up to hundreds of cameras, and many large cities around the world, including Beijing, London, New York and Seoul, have installed hundreds of thousands of cameras as part of their sur- veillance apparatus. Whilst large-scale camera network hardware is generally well-designed and well-installed, the development of intelligent video analysis software lags far behind.

Research on such intelligent video surveillance topics as person re-identification and human activity recognition has achieved great success in the past decade. Most of the algorithms involved are learning-based and assume that the joint distribution of the training data and label is similar to that in the testing phase. However, in large-scale camera networks, owing to different capturing environments, camera poses and other unpredictable conditions, this assumption is difficult to satisfy. The result is performance degradation, the degree of which depends on the level of joint distribution mismatch between the training data and testing data. A preliminary experiment on person re-identification showed that accuracy can be degraded by 13% in rank-10 accuracy (please refer to Section 2 for details). Therefore, the joint distri- bution mismatch problem must be resolved if existing video surveillance methods are to be deployed in large-scale camera networks.

Domain adaptation has been proved as a promising approach to solving the joint distribution mismatch problem. Existing algorithms can be roughly classified into three approaches: su- pervised, unsupervised and semi-supervised. Whilst different approaches are designed for different situations, each approach has its own assumptions/requirements, which are difficult to satisfy in most video surveillance applications in large-scale camera networks. Therefore, the performance of these algorithms may deteriorate dramatically, or they cannot be em- ployed directly. To overcome this problem, this project proposes a weakly supervised domain adaptation ap- proach, the basic idea of which is to employ the available information in large-scale camera networks to estimate weakly positive information (the weakly positive information refers to information about positive data rather than labels of that data). A new theory and algorithm allowing weakly supervised domain adaptation to learn and improve a classification model will then be developed on the basis of that estimated information. The theory and algorithm to be developed will be generic in nature, and applicable to video surveillance applications in large-scale camera networks.