An Inverted Residual based Lightweight Network for Object Detection in Sweeping Robots

Object detection is a core capability required by sweeping robots to conduct their cleanup jobs. Currently, deep learning-based object detection has achieved high accuracy; however, owing to the large number of backbone network layers, detection speed depends on the computing hardware, which hinders its application to real-time sweeping jobs. To address this problem, this paper proposes an accurate, fast, and lightweight You Only Look Once (YOLO) network for sweeping robots (SR). First, an improved online enhancement method is proposed to enrich the original dataset. Second, an inverted residual block based lightweight network is constructed for object recognition. Finally, an optimized spatial pyramid pooling method is added behind the backbone network to further improve network performance. Comparative experiments with many state-of-the-art methods are conducted, and the results reveal that the new model accuracy is 13.79% better than that of YOLOv4_tiny using a similar model size. Furthermore, its model size is only 1/9 that of YOLOv4 and achieves similar accuracy, demonstrating the advantages of the proposed model in terms of size, accuracy, and speed.

Automated summary

This paper proposes an accurate, fast, and lightweight YOLO network for sweeping robots. By using an improved online enhancement method, inverted residual blocks, and an optimized spatial pyramid pooling method, the model achieves good performance in terms of accuracy and speed.