This paper proposes an Efficient Perspective-n-Point (EPnP)-based You Only Look Once (YOLO) 3D object detection algorithm that aims to be applied to the navigation and localization system of autonomous robots when they work in outdoor environments. In the 3D object detection research area, although the classic PointNet series detection methods can learn shape and structure information directly from original point cloud data without complex preprocessing or feature extraction steps, the generalisation ability of this deep neural network model is limited by the training data. The EPNP based sensor fusion detection method proposed in this paper greatly reduces the dependence on the training data set due to its accuracy mostly reliant on the geometric transformation between camera and Lidar coordinate system. By combining YOLO this real-time 2D object detection method with EPnP pose estimation and Lidar data, this algorithm will be able to provide 3D object detection in real-time with less training data dependence. The key contribution is that the EPNP-based detection algorithm will first identify and classify targets in the 2D colour frames obtained by the depth camera, and use the absolute position relationship between the camera and the Lidar to get the camera pose relative to the Lidar coordinate system. Use the 2D pixel point coordinates that have also been transformed and the camera pose as input, apply EPnP in reverse to obtain the 3D point coordinates of the objects around the robots. The goal of this novel 3D object detection method through sensor fusion is to maintain a real-time speed while reducing the difficulty of training its depth model.