A Center Point-based Deep Learning Method for Monocular Depth Estimation

Engineering notes

Key topics: autonomous driving, bird's eye view, bev, depth estimation, monocular depth, carla, perception, prediction. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

In the domains of autonomous driving and roadside perception, the accurate location of vehicles and pedestrians is imperative for enhancing situational awareness and facilitating effective decision-making within vehicular network systems. However, the implementation of this capability through deep learning frequently necessitates the utilization of complex and resource-intensive multimodal datasets. To overcome this challenge, we develop a virtual dataset, designated as the Intersection Dataset, which includes extensive annotations of vehicles and pedestrians in various traffic scenarios. This dataset is created using the CARLA simulator. Furthermore, we introduce a new depth prediction method that directly regresses the depth of the target center relative to the camera through a defined mapping relationship. This innovative technique allows for accurate estimation of the Bird's Eye View (BEV) locations of vehicles and pedestrians in a scene. Our method is trained and evaluated on the Intersection Dataset and performs well in depth prediction across a wide range of spatial regions. To further validate its generalizability, we compare various backbone architectures to confirm the adaptability and robustness of the proposed framework.

Links and sources

• Official / arXiv page

Comments