Abstract: Aiming at the problem of insufficient extraction of 3D spatial features in 3D object detection models, we proposed an improved ICasA model. Firstly, we introduced focal sparse convolution to improve the 3D backbone network of CasA model, strengthening the extraction of information flow among non-empty voxel feature locations to obtain richer spatial features. Secondly, we constructed a multi-scale feature attention fusion module and embedded it into the 2D backbone network to enhance the model’s ability to handle multi-scale features. We adopted convolutions with different strides to deepen the structure of the 2D backbone network, improve the model’s ability to extract global features. We improved the attention fusion method to dynamically activate features of different scales and locations while preserving the original feature maps. The experimental results on the KITTI dataset show that compared with the CasA model, the ICasA model achieves 0.28 percentage points, 2.87 percentage points, and 1.03 percentage points increases in 3D average detection accuracy (AP@R40) for cars, pedestrians, and cyclists, respectively. Compared with existing advanced models, the ICasA model has more stable detection effect for all categories of objects, which helps to enhance the accuracy of 3D object detection.

Key words: 3D object detection, focal sparse convolution, multi-scale feature attention fusion, feature enhancement