Online objects localization using stereo camera

Feng Jiang, Wei Hua, Zhiyong Huang, Changhai Zha, Yechi Ma, Haiming Gao, Shun Zhang

Abstract: Autonomous driving technology has achieved great success in the past years. However, 3D perception from 2D images is a tough task cause it's an ill-posed problem for the visual-only perception. Deep learning requires large amount of training data and time consuming, while the perception results are often not generalized and difficult to be applied to other scenes. To solve this problem, we propose a binocular objects localization method based on traditional vision methods. We use a pre-trained model to detect objects in images. The feature points and their descriptors in the 2D bounding box are detected and matched, and the matched pairs are triangulated to obtain the 3D positions of the features. The point cloud filtering algorithm is used to pre-process the 3D features, and the CCP(contour central point) of the object is generated. Then the 3D bounding box model is built by the pre-defined size and the calculated orientation of each object. Due to the instability of single-frame detection, this paper uses several frames of detection results for modeling to obtain the predicted position of the object, and the results of the current frame are used as observation data. Kalman filter is used to obtain a more accurate estimated result. Finally, the proposed method is tested in a real scene and visualized by Baidu Apollo system, the results show that the method can detect objects effectively and real-time.