Video-based object tracking suffers from several limitations. The limited temporal resolution of 25 to 30 frames-per-second video data complicates the retrieval of fast moving objects when tracking them from one frame to the next. Furthermore video systems produce large amounts of data that puts great demands on any computational unit responsible for the data processing. A further challenge is in the necessity to operate under wide dynamic range lighting, which is prevalent in many application scenarios.

In contrast to traditional CCD or CMOS imagers that encode image irradiance and produce constant data volume at a fixed frame rate, irrespective of scene activity, event-based dynamic vision sensors contain an array of autonomous, self-signaling pixels which individually respond in real-time to relative changes in light intensity. Pixels that are not stimulated by a change in illumination do not produce output. Because there is no pixel readout clock, no initial time quantization takes place at the sensor, preserving the temporal information of the scene dynamics. Figure 1 shows a still image of a typical surveillance scene (left) and the sparse event data delivered a neuromorphic dynamic vision sensor, containing all dynamic scene information relevant for efficient and robust real-time object tracking (right).

An event-based tracking algorithm which is inspired by the mean-shift approach works on the basis of instantaneous clustering of the asynchronous pixel events, and continuous-time tracking of active clusters. The algorithm processes each event as it is received without buffering and in real-time, hence only a very limited amount of memory and computing power is required.

Based on this technology, compact, low-power and cost-effective embedded systems for real-time object tracking can be realized.