DenseLightNet: A Light-Weight Vehicle Detection Network for Autonomous Driving
| Autor: | Qiwei Ding, Qin Zou, Long Chen, Lingxi Li, Zhaotang Chen |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Computer science
020208 electrical & electronic engineering Detector Feature extraction Real-time computing 02 engineering and technology Convolutional neural network Object detection Power (physics) Convolution Network planning and design Control and Systems Engineering Range (aeronautics) 0202 electrical engineering electronic engineering information engineering Electrical and Electronic Engineering |
| Zdroj: | IEEE Transactions on Industrial Electronics. 67:10600-10609 |
| ISSN: | 1557-9948 0278-0046 |
| DOI: | 10.1109/tie.2019.2962413 |
| Popis: | In recent years, vehicle detectors built on deep convolutional neural network (DCNN) have been widely used in autonomous driving. Under the complex conditions of road traffic, the detector is expected to run in high speed and high accuracy. However, due to the limited computing power and storage space on the autonomous vehicle, the deployment of advanced DCNN detectors is often restricted. The design of lightweight and powerful detectors is in a great desire. Recently, group convolution, as a novel convolution algorithm, has been proposed to reduce the floating-point operations and make the detection network lighter and faster. However, in practice, it is found that the increase of group number does not always boost the detection speed, but sometimes leads to the performance degradation. In addition, the existing guidelines for network design do not indicate how to choose the group number in the group convolution in order to maximize the overall detection speed. To this end, in this paper, we propose three new guidelines to determine the valid range of the group number, and design a lightweight detection network—DenseLightNet—based on these new design criteria. The proposed detector runs at a speed of three times faster than the current real-time detector YoloV3, while holding a much smaller model size. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|