Abstract
Autonomous driving research often prioritizes accuracy at the cost of high computational overhead, making many state-of-the-art models impractical for mass-market embedded systems. We present RAD (Realtime and Accurate Detector), a 3D object detection framework specifically optimized for low-power hardware. By utilizing a 3D binary occupancy grid encoding and a streamlined convolutional architecture, RAD achieves nearly double the inference speed of previous industry standards like PointPillars while maintaining comparable detection performance on the KITTI benchmark. Our approach ensures density-independent processing, making it robust for various LiDAR configurations in real-world Advanced Driver Assistance Systems (ADAS).
Introduction
While high-end GPUs are standard in research environments, production vehicles typically utilize low-power embedded systems. These systems must process massive LiDAR point clouds in real-time to ensure safety. Existing methods often struggle with computational complexity—relying on 3D sparse convolutions that scale poorly with point cloud density—or suffer from pre-processing bottlenecks like "pillarization."
RAD addresses these challenges by focusing on a detector that is both density-independent and fast enough to operate comfortably on embedded boards, such as the NVIDIA Jetson AGX Xavier, without sacrificing the accuracy required for reliable object detection.
Methodology
The innovation of RAD lies in its simplified input encoding and efficient neural network architecture.
Fast Input Encoding
Instead of learning complex features for voxels, RAD employs a 3D binary occupancy grid. A cell is marked as 1 if it contains at least one point and 0 otherwise. Given that LiDAR data is approximately 99.88% sparse, the framework uses a "Dictionary of Keys" (DOK) representation. This allows for the efficient transfer of only non-zero data from the CPU to the GPU, significantly reducing data transfer overhead.
Streamlined Architecture
The RAD architecture consists of:
- Encoder: Three blocks of standard $3 \times 3$ convolutions. Spatial dimensions are downsampled quickly using a stride of 2 in the initial convolution of each block.
- Decoder: A lightweight structure that concatenates feature maps from multiple scales. Inference is further optimized by rearranging upsampling layers to maintain network "width" without increasing computational cost.
- Detection Heads: Three specialized heads handle classification (object presence), regression (localization and dimensions), and direction (orientation).
Experimental Results
RAD was evaluated on the KITTI Vision Benchmark Suite, focusing on the Car and Cyclist categories using the standard training/validation split.
Performance and Speed
RAD sets a new benchmark for speed-to-accuracy efficiency. On a high-end RTX 2080Ti, RAD achieves 84.46 FPS, doubling the performance of PointPillars (42.03 FPS). More importantly, on the target embedded hardware (Jetson AGX Xavier), RAD reaches 10.91 FPS, a 77% improvement over PointPillars.
Accuracy Analysis
In the KITTI "Easy" Car category (3D metric), RAD achieved 88.70% AP, outperforming PointPillars' 86.00%. The model showed particular strength in cyclist detection across all difficulty levels. The authors noted that while standard IoU thresholds (0.7) are strict, RAD remains highly effective at an IoU of 0.5, which is sufficient for critical ADAS functions like collision warnings.
Discussion and Future Work
Analysis of regression errors indicates that small inaccuracies in depth ($z$) and length ($l$) are the primary factors limiting higher AP scores at strict IoU thresholds. Future iterations will focus on stabilizing these regression heads. Additionally, while the model is highly efficient, it shows sensitivity to data augmentation; specifically, the vertical alignment of augmented samples must be precisely corrected against the ground plane to maintain 3D accuracy.
In conclusion, RAD demonstrates that a simplified, occupancy-based approach can provide the necessary speed for embedded ADAS applications while remaining competitive with far more complex, resource-intensive models.
Summary updated