Fast and Efficient Non-Contact Ball Detector for Picking Robots

Abstract

Object detectors based on deep learning requires high-performance computing and large run-time memory footprint to maintain good detection performance. They bring high computation overhead and power consumption to on-board embedded devices of non-contact ball picking robot. Furthermore, it is difficult to deploy on the machine because the model size is so big. The accuracy of the existing simplified detectors deployed on embedded devices cannot meet the requirements of practical applications. Therefore, how to reduce floating point operations (FLOPs) and the size of model without notably sacrificing detection precision becomes an urgent problem to be solved. To solve this problem, a shuttle residual block which is more efficient network unit based on depthwise separable convolution was proposed. And we designed a non-contact ball object detector for picking robots, which is shallower than YOLOv3 and has narrower structure. We evaluate the proposed method on non-contact Ball dataset and compelling results are achieved by the proposed method. Compared with YOLOv3, the proposed method reduces FLOPs by 86.2%, declines parameter size by 89.5%. Overall, the proposed method achieves comparable detection accuracy than YOLOv3, and its speed is 2.2 times faster than YOLOv3.

Keywords

• Depthwise separable convolution
• picking robots
• embedded device
• deep learning