OccMALA: An Efficient 3D Semantic Occupancy Prediction Method with Linear Complexity for Perception on Autonomous Ground Platforms

Engineering notes

Extensive experiments on the SemanticKITTI benchmark demonstrate that OccMALA achieves a 0.4% improvement in mean Intersection over Union (mIoU) compared to the current mainstream LiDAR-based method, SSC-RS. More importantly, it significantly boosts the inference speed by 40.1%.

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

For autonomous ground platforms operating in scenarios such as unmanned inspection and delivery, efficient and accurate perception of the surrounding environment's geometric structure and semantic information is essential. Compared to general 3D object detection methods, 3D semantic occupancy prediction excels at modeling irregular objects and providing precise semantic information, thus emerging as a mainstream approach in autonomous driving perception. However, most existing 3D semantic occupancy prediction methods are constructed based on 3D CNNs or Transformers, which impose high computational and memory demands. This makes them difficult to deploy on resource-constrained small autonomous ground platforms and limits their global modeling capabilities. To address the issues of high computational complexity and poor global modeling in current methods, we propose a lightweight and efficient occupancy prediction network named OccMALA. The contributions are threefold: First, we explicitly decouple the learning processes of semantic segmentation and scene completion, and train them with multi-scale supervision separately. Second, to tackle the high computational cost and memory consumption associated with voxel feature extraction, we project voxel features into the Bird's-Eye-View (BEV) space to reduce computational overhead. Third, for efficient feature extraction and fusion of semantic segmentation and scene completion features, we introduce the BEV-MALA module to enhance BEV features and propose the U-MALA architecture for adaptive multi-scale BEV feature fusion. Extensive experiments on the SemanticKITTI benchmark demonstrate that OccMALA achieves a 0.4% improvement in mean Intersection over Union (mIoU) compared to the current mainstream LiDAR-based method, SSC-RS. More importantly, it significantly boosts the inference speed by 40.1%. Furthermore, we deploy OccMALA on an autonomous ground platform equipped with the NVIDIA Jetson Xavier NX as its onboard computer, where it achieves a real-time inference time of 53.8 ms, representing a 35.3% improvement over SSC-RS, while maintaining a lower GPU memory consumption.

Links and sources

• Official / arXiv page
• PDF from original source

Comments