Year: 2022
Type: conference-paper
Abstract: Lidars and cameras are critical sensors which offer complementary information for scene perception and trajectory planning tasks in autonomous driving. In this work, we propose an attention-based framework to integrate representations of information from the two sensors and complete trajectory planning tasks using an end-to-end learning-based approach. To fuse 3D point cloud with RGB image in the same dimension, we also utilize a gradient transformation operation to convert point cloud into grey images. For the trajectory planning task, we use an auto-regressive network consisting of GRUs to predict future trajectory of ego-vehicle. The experiment results show that the proposed framework outperforms the models with single sensor input and the model without cross attention. We demonstrate the efficacy of our approach in the CARLA driving simulator.
Author: Yilin Zhang
Institution: National University of Defense Technology
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4281864424