3D Object Detection Based on Long and Short Focal Length Cameras

3D object detection is an important technology for environment perception of autonomous vehicle. The 3 D object detection method based on monocular images cannot achieve both tele-view and wide-view simultaneously. To solve this problem, this paper adopts the combination of long and short focal length cameras. The short focal length camera produces wide-view images, and the long focal length camera produces the tele-view images. Firstly, the DLA-34 feature extraction network is used to extract features from both long-focal and short-focal images. Then, keypoint detection and 3D box parameter regression are performed on the image features to obtain detection results for both types of images. The detection results of long-focal images are projected onto short-focal images using image matching methods, and the detection results of the same object are fused to achieve 3D object detection in both tele-view and wide-view. Evaluated on the self-built dataset, our method has an average accuracy of $56.31 \%$, which shows a significant improvement.