Object Detection and Classification in FWMAVs for Smart Pollination

Object Detection and Classification algorithms are necessary for Smart Pollination by FWMAVs(Flapping-Wing Micro Aerial Vehicles). The FWMAVs would need to recognise flowers and navigate towards them. Due to the extremely low limitations on size, power consumption and weight, heavy machine learning models could not be developed to run on small microcontroller chips due to heavy resource consumption and low speed. In this research, we try to overcome this problem by having the FWMAV stream the video in real-time to a server and have the object detection algorithms run online. We implement this in two ways and compare them. In the first one, we directly upload the stream without processing, and in the second, we compress the data before uploading. We compare both methods on the basis of accuracy and performance. We conclude that the former is ideal for crispier data and an environment with lower latency. The latter is advisable for those areas where network has high latency and less data transfer is feasible.