Abstract
Lightweight UAVs equipped with deep learning models have become a trend, which can be deployed for automatic navigation in a wide range of civilian and military missions. However, real-time applications usually need to process a large amount of image data, which leads to a very large computational complexity and storage consumption, and restricts its deployment on resource-constrained embedded edge devices. To reduce the computing requirements and storage occupancy of the neural network model, we proposed the ensemble binarized DroNet (EBDN) model, which implemented the reconstructed DroNet with the binarized and ensemble learning method, so that the model size of DroNet was effectively compressed, and ensemble learning method was used to overcome the defect of the poor performance of the low-precision network. Compared to the original DroNet, EBDN saves more than 7 times of memory footprint with similar model accuracy. Meanwhile, we also proposed a novel and high-efficiency hardware architecture to realize the EBDN on the chip (EBDNoC) system, which perfectly realizes the mapping of an algorithm model to hardware architecture. Compared to other solutions, the proposed architecture achieves about 10.21 GOP/s/kLUTs resource efficiency and 208.1 GOP/s/W energy efficiency, while also providing a good trade-off between model performance and resource utilization.
Highlights
Over the past few years, deep convolutional neural networks (DCNNs) have been extensively studied and applied due to their excellent performance [1], especially in the challenging field of computer vision, such as image classification [2,3], object detection [4,5], and instance segmentation [6,7]
Compared to the original full-precision model, the ensemble binarized DroNet (EBDN) model achieves a similar accuracy on Udacity and Collision datasets with much less memory footprint, which dramatically reduces the power consumption of the hardware
This paper proposed the EBDN model, which is a perception-control integrated deep binarized DroNet model with a high energy efficiency and low hardware resource cost for visual navigation of unmanned aerial vehicles (UAVs) application
Summary
Over the past few years, deep convolutional neural networks (DCNNs) have been extensively studied and applied due to their excellent performance [1], especially in the challenging field of computer vision, such as image classification [2,3], object detection [4,5], and instance segmentation [6,7]. This paper presents a hardware architecture design of the ensemble binarized DroNet (EBDN) model for realtime UAV autonomous navigation engine, which integrates the capacity of perception and control. It is the novel approach developed for the deep neural network acceleration via model compression for efficient FPGA implementation. We propose an EBDN model, which overcomes the accuracy bottleneck of a single network by integrating several binarized DroNets with ultra-low memory footprint, and can dramatically speed up the inference process of the model when compared to full precision neural networks.
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
