A High-speed Event-based Object Tracking with Heterogeneous Computing Hardware

Object tracking is the core of tasks such as environment perception and vision navigation. Event cameras are neuromorphic vision sensor that encode relative changes in light intensity rather than absolute intensity. The sensitive nature of event cameras to dynamics is well suited for object tracking. Meanwhile, high-speed hardware implementation of event-based algorithms has always been a major problem in the field of event-based vision, because traditional data processing hardware is not suitable for event stream processing. Unfortunately, neuromorphic computing hardware suitable for processing events is still at the dawn, due to the mechanisms for training neuromorphic models that are not yet clear. We aim to implement this algorithm in hardware to further demonstrate the possibilities of event-based vision in practical applications. However, the existing solutions are mainly based on homogeneous hardware, which does not take full advantage of the sparse and asynchronous nature of events. In order to solve these problems, a heterogeneous computing architecture and hardware designs are proposed for high-speed implementation of event-based object tracking. The experimental result shows that the computing speed in our proposed heterogeneous computing hardware implementation is 1.67 times faster than that on homogeneous computing hardware. Meanwhile, a more efficient power consumption performance is obtained.