NeuroVI-based new datasets and space attention network for the recognition and falling detection of delivery packages

Abstract

With the popularity of online-shopping, more and more delivery packages have led to stacking at sorting centers. Robotic detection can improve sorting efficiency. Standard datasets in computer vision are crucial for visual detection. A neuromorphic vision (NeuroVI) camera is a bio-inspired camera that can capture dynamic changes of pixels in the environment and filter out redundant background information with low latency. NeuroVI records pixel changes in the environment with the output of event-points, which are very suitable for the detection of delivery packages. However, there is currently no logistics dataset with the sensor, which limits its application prospects. This paper encodes the events stream of delivery packages, and converts the event-points into frame image datasets for recognition. Considering the falling risk during the packages' transportation on the sorting belt, another falling dataset is made for the first time. Finally, we combine different encoding images to enhance the feature-extraction on the YOLO network. The comparative results show that the new datasets and image-confusing network can improve the detection accuracy with the new NeuroVI.