MOCA: Monitoring of Crowd Activities

The MOCA project will develop a network-enabled DSP+FPGA-enhanced image-intensified embedded system for self-contained low-power monitoring of crowds in open and in complex urban terrain. The Phase I effort will develop new algorithms explicitly for crowd management, rather than building on the traditional intelligent video surveillance algorithms that are focused on isolated targets in motion. Phase I will also address re-identification of the same individual across sensors and time. The sensor, building on past projects, will be able to support an image-based detection and report classes of targets (e.g., reporting a dismounted human, van or pickup) as well as individual face-based biometrics. While it can support them, the system design departs from the more traditional background-subtraction-based approaches that have been shown difficult to extend to complex environments with considerable moving clutter and dense crowds. The proposal's approach combines unique crowd-oriented real-time optic-flow analysis with a hypothesize/recognize and track approach for larger scale motion and sensor-to-sensor handoff. The effort will analyze the use of DSP+ FPGAs to support real-time computation of features of flow and for recognition/detection at low power. Without the massive compute power of the 1.3Million gate FPGA + a 600Mhz DSP, processing the mega-pixel images would not be possible in a low cost 7w system weighting under 2kg. The combination of recognition and flow-based processing, both enabled by the FPGA acceleration, will allow the system to work on active PTZs and with dense moving crowds. The feature-enriched tracking will support tracking through occlusion, tracking from moving sensors, tracking in clutter, and sensor-to-sensor handoff.