Research on Small Target Detection Based on Optimised PointPillars Backbone Algorithm

ABSTRACT

Aiming at the problem of low detection accuracy of the PointPillars algorithm for small targets, an optimised target detection algorithm based on the PointPillars backbone network is proposed. Firstly, the 2D convolution module of PointPillars backbone network is improved based on the depth separable volume, which improves the correlation between the model channels and ensures the detection accuracy, and also makes the model lighter to a certain extent; then, the 2D convolution module of PointPillars backbone network is improved based on the deformable convolution, which makes the model have a more accurate prediction of the target position and improves the detection accuracy. Finally, the squeeze‐and‐excitation and coordinate attention mechanisms are incorporated into the model to enhance the detection accuracy further. The results of the experiment on the KITTI dataset show that the improved model has a higher detection accuracy for small targets compared with PointPillars. Among all compared models, Df‑pointpillars‑CA(16) yielded the largest gains on the four metrics, improving by up to 3.63%, 1.90%, 0.93% and 2.17% for the pedestrian category, and by up to 4.33%, 5.85%, 4.80% and 3.46% for the Cyclist category, respectively.