FPGA Accelerated CNN Inference and Online Retraining for UAVs with Minimal Power and Performance Overhead

Unmanned Aerial Vehicles (UAVs) are quickly becoming a very important component in many applications, such as search and rescue. In many of these applications, Deep Neural Networks (DNNs) are used, especially for obstacle detection and avoidance. Furthermore, online retraining may be required for adaptation to new environments. This paper shows how to enable UAVs equipped with an FPGA-based controller to achieve the required accuracy, whether in the inference phase or in the retraining phase, with minimal impact on power and performance. Dynamic Function eXchange (DFX) is used to download the retraining module bitstream utilizing Posit16 multipliers when the UAV navigates a new environment. Otherwise, a Posit8 multiplier (lower in precision) based implementation is used for the inference procedure that requires less precision than the retraining procedure. An implementation on the EP4CE22F17C6 FPGA is presented to show that a Posit16 multiplier has very close power utilization and performance to that of a Posit8 multiplier.