Video deblurring and flow-guided feature aggregation for obstacle detection in agricultural videos

Autonomous agricultural vehicles are increasingly common on farms, where they can replace humans in tasks such as irrigation, harvesting, and weeding, reducing labor costs. Real-time obstacle avoidance is a prerequisite for their work. At present, vehicles equipped with vision sensors cannot perform end-to-end video object detection, and their accuracy is also affected by motion blur. We propose a novel agricultural obstacle detection method based on RNN and flow-guided feature aggregation, combining video deblurring and object detection tasks for joint optimization. In addition, to make full use of the region proposals, a region shared strategy is proposed to improve the efficiency of video deblurring. The proposed method can solve the common motion blur problem in agricultural video and is expected to be suitable for all kinds of obstacle detection tasks in agricultural scenes. We experimented with this method on the FieldSAFE and GOPRO datasets. Our method provides better detection performance and is computationally less costly than other methods according to experimental results.