Abstract: A real-time and robust 3D dynamic object perception module is a key part of autonomous driving system. This paper fuse monocular camera and LiDAR to detect 3D objects. Firstly, we use a convolutional neural network (CNN) to detect 2D bounding boxes in the image and generate 3D frustum regions of interest (ROI) according to the geometric projection relation between lidar and camera. And then, we cluster the point cloud in the frustum ROI and fit the 3D bounding box of the object. After detecting 3D objects, we reidentify the objects between adjacent frames by appearance features and Hungarian algorithm, and then propose a tracker management model based on a quad-state machine. Finally, a novel prediction model is proposed, which leverages lane lines to constrain vehicle trajectories. The experimental results demonstrate that our algorithm is both effective and efficient. The whole algorithm only takes approximately 25 milliseconds, which meets the requirements of real-time.