ACHIEVING COMMERICALLY-VIABLE ROBUST VIDEO-BASED UNSTRUCTURED SCENE UNDERSTANDING SYSTEMS

Today's commercial video surveillance and monitoring (VSAM) systems are capable of producing enormous video streams that are over-whelming for human operators to review. But with automated intelligent interpretation of the imagery we can track and classifymoving objects of interest, collect spatial and temporal motion statistics, and detect anomalous behaviors. We propose to transition image understanding (IU) technologies developed at MIT's Artificial Intelligence Lab into VSAM systems under development byALPHATECH, Inc. to achieve robust scene interpretation capabilities. Our plan to accomplish this technology transition consists of leveraging high performance motion segmentation, tracking, and classification techniques, identifying their benefits andlimitations, developing application-oriented end-to-end modular system designs, and developing and executing VSAM performance characterization methods. We target unstructured outdoor and indoor monitoring applications in which complete access control isnot feasible. Video-based scene interpretation provides a low-cost information-rich data source that allows us to dynamically characterize activities in these scenes according to both the spatial properties and temporal properties of objects in the scene.By using robust and adaptable video processing technology we achieve effective scene monitoring performance across a wide range of viewing conditions. We are developing products for (1) security applications, detecting suspicious activity by identifyingtraffic pattern anomalies or by measuring loitering statistics in sensitive regions of interest; (2) traffic flow monitoring, providing live traffic flow feeds, comprising speed and count statistics, for municipal government agencies; and (3) activitycharacterization, defining average traffic patterns as a function of location and time for characterizing shopper preferences and travel bottlenecks.