An Attention Mechanism Based AVOD Network for 3D Vehicle Detection

Publisher:
Institute of Electrical and Electronics Engineers (IEEE)
Publication Type:
Journal Article
Citation:
IEEE Transactions on Intelligent Vehicles, 2023, PP, (99), pp. 1-13
Issue Date:
2023-01-01
Filename Description Size
An Attention Mechanism Based AVOD Network for 3D Vehicle Detection_OPUS.pdfAccepted version9.38 MB
Adobe PDF
Full metadata record
With the continuous advancement of autonomous driving technology, 3D vehicle detection has become of widespread interest. The traditional aggregate view object detection (AVOD) framework has achieved some good results in 3D vehicle detection tasks. However, the complexity of the 3D vehicle detection scenario makes the current detection methods still not meet the actual requirements. To enhance the detection accuracy of 3D vehicle targets, we propose to equip an attention mechanism to improve the representation capability of feature maps, thereby further increasing the precision of 3D vehicle detection. Specifically, we have added the channel attention ECANet, spatial attention SANet, and mixed attention ECANet+SANet respectively into the image-based feature pyramid network of the AVOD detection framework, which can enhance the feature maps representation and improve the detection accuracy observably. The improved AVOD network is verified using the KITTI dataset. By showing the detection results of these attention mechanisms, it is found that the feature pyramid networks (FPN) module in the AVOD network based on Image has the best performance when integrating a mixed attention mechanism. In comparison to the original AVOD network, the detection results on the average precision index of the proposed method have improved by 2.29%, 2.81%, and 1.32% in the three indexes of simple, medium, and difficult, respectively. Extensive experiments have confirmed the practicality and efficacy of the AVOD network to equip the attention mechanisms for 3D vehicle detection.
Please use this identifier to cite or link to this item: