Data Mining Based Learning Algorithms for Semi-Supervised Object Identification and Tracking (ACM)
Sensor exploitation (SE) is the crucial step in surveillance applications such as airport security and search and rescue operations. It allows localization and identification of movement in urban settings and can significantly boost knowledge gathering, interpretation and action. Data mining techniques offer the promise of precise and accurate knowledge acquisition techniques in high-dimensional data domains (and diminishing the “curse of dimensionality” prevalent in such datasets), coupled by algorithmic design in feature extraction, discriminative ranking, feature fusion and supervised learning (classification). Consequently, data mining techniques and algorithms can be used to refine and process captured data and to detect, recognize, classify, and track objects with predictable high degrees of specificity and sensitivity.
Automatic object detection and tracking algorithms face several obstacles, such as large and incomplete datasets, ill-defined regions of interest (ROIs), variable scalability, lack of compactness, angular regions, partial occlusions, environmental variables, and unknown potential object classes, which work against their ability to achieve accurate real-time results. Methods must produce fast and accurate results by streamlining image processing, data compression and reduction, feature extraction, classification, and tracking algorithms. Data mining techniques can sufficiently address these challenges by implementing efficient and accurate dimensionality reduction with feature extraction to refine incomplete (ill-partitioning) data-space and addressing challenges related to object classification, intra-class variability, and inter-class dependencies.
A series of methods have been developed to combat many of the challenges for the purpose of creating a sensor exploitation and tracking framework for real time image sensor inputs. The framework has been broken down into a series of sub-routines, which work in both series and parallel to accomplish tasks such as image pre-processing, data reduction, segmentation, object detection, tracking, and classification. These methods can be implemented either independently or together to form a synergistic solution to object detection and tracking.
The main contributions to the SE field include novel feature extraction methods for highly discriminative object detection, classification, and tracking. Also, a new supervised classification scheme is presented for detecting objects in urban environments. This scheme incorporates both novel features and non-maximal suppression to reduce false alarms, which can be abundant in cluttered environments such as cities. Lastly, a performance evaluation of Graphical Processing Unit (GPU) implementations of the subtask algorithms is presented, which provides insight into speed-up gains throughout the SE framework to improve design for real time applications.
The overall framework provides a comprehensive SE system, which can be tailored for integration into a layered sensing scheme to provide the war fighter with automated assistance and support. As more sensor technology and integration continues to advance, this SE framework can provide faster and more accurate decision support for both intelligence and civilian applications.
Paper available at ACM.