Optics and Precision Engineering, Volume. 31, Issue 19, 2910(2023)
Improved PointPillar point cloud object detection based on feature fusion
A point cloud object detection network, Pillar-FFNet, is proposed by introducing a multiscale feature fusion strategy and an attention mechanism to address the ineffectiveness of PointPillar in detecting small sparse objects in point clouds in autonomous driving road scenarios. First, a backbone network based on a residual structure is designed for feature extraction in the network. Second, a simple and effective multiscale feature fusion strategy is designed to address the problem that the feature maps fed into the detection head do not make full use of the semantic information of high-level features and the spatial information of low-level features. Finally, a convolutional attention mechanism is proposed to treat information redundancy in the feature maps extracted using the backbone network. To validate the performance of the proposed algorithm, experiments are conducted on the KITTI and DAIR-V2X-I datasets. The results show that the proposed algorithm achieves maximum average accuracy improvements of 0.84%, 2.13%, and 4.02% for cars, pedestrians, and cyclists, respectively, on the KITTI dataset and maximum average accuracy improvements of 0.33%, 2.09%, and 4.71% for cars, pedestrians, and cyclists, respectively, on the DAIR-V2X-I dataset compared with the PointPillar results. Experimental results demonstrate the effectiveness of the proposed method for the detection of sparse small objects in point clouds.
Get Citation
Copy Citation Text
Yong ZHANG, Zhiguang SHI, Qi SHEN, Yan ZHANG, Yu ZHANG. Improved PointPillar point cloud object detection based on feature fusion[J]. Optics and Precision Engineering, 2023, 31(19): 2910
Category:
Received: Mar. 30, 2023
Accepted: --
Published Online: Mar. 18, 2024
The Author Email: Zhiguang SHI (szgstone75@sina.com)