Effect of Post-Processing Methods on CPU-based 3D Object Detection from Camera and LiDAR Data

This paper investigates the effect of several post-processing techniques on the accuracy and consistency of a lightweight 3D object detection pipeline using camera and LiDAR data fusion. The baseline camera-LiDAR fusion solution, which we extend in this work, integrates a YOLO-based segmentation with LiDAR point data, without further use of neural networks. This approach focuses on inference suitable for resource-constrained environments. To enhance detection performance, several improvements were introduced: bounding box height calibration using segmentation masks and LiDAR depth data, merging overlapping detections for cyclists, filtering background points from bicycles, and optimizing the center position and orientation of bounding boxes for elongated objects through convex hull analysis and rotation optimization. The evaluation conducted on the View of Delft (VoD) dataset confirmed that these enhancements improved the mean Average Precision (mAP) and Intersection over Union (IoU) for object classes car, cyclist, pedestrian, motorcycle and truck.