Real-time Object Detection Algorithm for Apple Picking Robots Based on RT-DETR-Faster
In order to solve the problems of small targets and high real-time requirements in apple picking,this paper proposes a target detection method for picking robots based on RT-DETR,named RT-DETR-Faster.Firstly,the paper adopts FasterNet partial convolution to replace the traditional convolution of the backbone network,which effectively improves the model's computation speed.Secondly,the paper uses an improved cascaded attention encoder to replace the original encoder,which makes the network more focused on the target area.Finally,the paper introduces Faster_Rep feature fusion module,which preserves more effective features and reduces the computation cost.Conduct-ed experiments on real orchard images to evaluate the effectiveness of proposed algorithm.The results show that our algorithm achieves a 34%increase in FPS,reaching 47.9 fps,and a higher accuracy compared to the original RT-DE-TR algorithm,making it suitable for real-time object detection tasks of apple picking robots.
deep learningorchard pickingTransformerattention mechanismRT-DETR
NETL
NSTL
万方数据
