3D object detection from LiDAR is a crucial task in autonomous driving and mobile robotics, providing precise 3D bounding boxes of obstacles. State-of-the-art detectors achieve high accuracy but at the cost of processing entire point clouds with heavy neural networks, which makes real-time deployment on resource-constrained edge devices and low-power robots challenging. In this paper, we propose an adaptive cropping strategy for real-time processing of LiDAR data, which significantly reduces computation by focusing only on regions of interest across frames while maintaining detection accuracy. Our method leverages bounding box predictions and vehicle state estimates from each frame to crop the subsequent frame’s point cloud to likely object regions, thereby significantly reducing the data volume processed. We integrate this pipeline with modern one-stage and two-stage LiDAR detectors and evaluate on both a custom roadside LiDAR dataset and the public nuScenes dataset. Experimental results show that our approach can process as little as 50% of the points with only a minor mAP decrease for vehicles and pedestrians. On an NVIDIA Jetson AGX Orin edge device, the pipeline significantly improves inference speed, roughly doubling the system FPS. These results demonstrate the feasibility of efficient real-time 3D perception on low-power systems like mobile robots and edge devices.