A low power smart camera for video surveillance and forensic applications

Abstract

We describe a custom, low-power surveillance camera for the control of isolated, peripheral areas, that are often chosen for illegal activities, like traffic of drugs, weapons and cigarettes. The camera embeds on chip an image processing algorithm for real-time event detection, exploiting a dynamic background subtraction approach to identify the so-called hot-pixels, i.e. pixels undergone an intensity variation, and possibly corresponding to an abnormal action. Every time a group of hot-pixels is detected, the camera generates an alarm that can be sent to an external processor for further processing (e.g. scene classification) or to a surveillance officer to plan an intervention. The related frames are saved to be used as forensic evidence in a court. The presented vision sensor combines the image array with a bench of processors, frame buffer, timing control block and digital interface. Low-power performance is obtained through custom chip design techniques, combined with integrated image processing and different sensor operating modes to minimize the sensor average power consumption. The vision sensor performance has been evaluated by two European LEAs, which expressed a high level of interest in such a system and a positive vote about its event detection performance.