RADE-Net: Robust Attention Network for Radar-Only Object Detection in Adverse Weather

The paper proposes RADE-Net, a lightweight radar-only object detection model that utilizes a novel 3D projection of 4D RADE tensors to preserve rich Doppler and elevation features while significantly reducing data size, achieving superior performance over existing radar and lidar methods in adverse weather conditions.

The Gist

Imagine you are driving a car in a heavy fog, pouring rain, or a blizzard. Your eyes (cameras) are useless because they can't see through the whiteout. Your high-tech laser scanner (LiDAR) is like a sensitive microphone that gets overwhelmed by the noise of the rain and snow, creating a chaotic mess of data.

But your car's Radar? It's like a bat using echolocation. It doesn't care about the fog; it just bounces sound waves off objects and listens for the echo. It sees the car in front of you clearly, even when you can't.

However, there's a problem. While Radar is great at seeing through bad weather, the raw data it sends back is like a giant, messy, 4-dimensional ocean of numbers. It's too heavy for a car's computer to process quickly, and most existing AI models are too simple to understand the nuances of this data. They often throw away the most interesting parts of the signal to save space, losing valuable clues about how fast an object is moving or how tall it is.

Enter RADE-Net, a new "brain" for self-driving cars designed to solve this exact problem.

The Core Idea: The "Smart Squeeze"

Think of the raw Radar data as a massive, 4D block of Jell-O containing Range (distance), Azimuth (angle), Doppler (speed), and Elevation (height).

Most previous methods tried to slice this Jell-O into flat 2D pancakes to make it easier to eat. But in doing so, they lost the flavor (the speed and height data). Other methods tried to eat the whole block but got a stomach ache (too much memory, too slow).

RADE-Net uses a "Smart Squeeze." It takes that giant 4D block and folds it into a compact 3D sandwich.

• It keeps the Distance and Angle (where the object is).
• It tucks the Speed and Height data right into the layers of the sandwich.
• The Result: It shrinks the data size by 92% (like turning a giant suitcase into a carry-on bag) without throwing away a single crumb of information. This makes the AI incredibly fast and efficient.

The "Detective" Architecture

Once the data is squeezed into this 3D sandwich, it goes through a special detective process called RADE-Net:

1. The Magnifying Glass (Attention Mechanism): Imagine a detective looking at a crime scene. They don't look at everything equally; they focus on the clues that matter. RADE-Net has a built-in "Attention Module" that acts like a magnifying glass. It highlights the most important parts of the Radar signal (like a car's speed or a pedestrian's height) and ignores the background noise.
2. The Two-Step Dance (Decoupled Heads): Instead of trying to guess the object's location and shape all at once, RADE-Net does it in two smooth steps:
   • Step 1: It finds the exact center point of the object on a map (like dropping a pin on Google Maps).
   • Step 2: It draws a 3D box around that pin, figuring out how long, wide, tall, and which way the object is facing.
3. The Translator: The Radar "speaks" in polar coordinates (distance and angle), but the car needs to drive in Cartesian coordinates (X, Y, Z on a grid). RADE-Net is a master translator, instantly converting the Radar's "map" into a real-world 3D view.

Why It's a Game Changer

The researchers tested this new system on the K-Radar dataset, which is basically a massive library of driving scenes in terrible weather (fog, snow, rain).

• Beating the Baseline: Compared to the previous best Radar-only AI, RADE-Net improved detection accuracy by 16.7%. That's like going from a blurry security camera to a crystal-clear HD feed.
• Beating the "Expensive" Sensors: In normal weather, expensive LiDAR and Cameras usually win. But in a foggy scenario? RADE-Net didn't just hold its own; it crushed the LiDAR systems by 32%. It proved that when the weather turns nasty, the humble Radar, powered by this new AI, is the king of the road.
• Speed: Because it's so lightweight, it can run on the small, low-power computers inside a car without overheating.

The Bottom Line

RADE-Net is like giving a self-driving car a pair of super-vision glasses that work perfectly in a storm. It takes the messy, heavy data from a Radar sensor, cleans it up, and uses smart attention to find cars, trucks, and even pedestrians, even when the world is white with snow or gray with fog. It proves that you don't need the most expensive sensors to drive safely; you just need the right way to look at the data you already have.

Technical Summary

1. Problem Statement

Automotive perception systems face significant challenges in adverse weather conditions (fog, rain, snow). While optical sensors like Cameras and LiDAR offer high resolution, they suffer from performance degradation or signal blockage in these environments. Radar sensors, operating at longer wavelengths, penetrate adverse weather effectively but present their own challenges:

• Data Sparsity & Information Loss: Traditional Radar processing converts 4D tensors (Range, Azimuth, Doppler, Elevation) into sparse point clouds via thresholding (e.g., CFAR). This process discards rich structural and velocity information, leading to poor detection of vulnerable road users (pedestrians, cyclists).
• Computational Burden: Utilizing full 4D Radar tensors directly for deep learning is memory-intensive, making real-time training and inference difficult on embedded hardware.
• Feature Disconnection: Existing 2D projection methods often lose the inherent relationship between positional bins (Range-Azimuth) and velocity/elevation features. Conversely, methods using sparse tensors or full 3D CNNs often incur excessive computational costs.

2. Methodology

The authors propose RADE-Net, a lightweight deep learning framework designed to process 4D Radar tensors efficiently while preserving critical Doppler and Elevation features.

A. Tensor Projection Module (TPM)

To address memory constraints without losing information, the authors introduce a 3D projection method:

• Mechanism: The full 4D FFT-processed tensor ($R \times A \times D \times E$) is split into two projections: Range-Azimuth-Doppler (RAD) and Range-Azimuth-Elevation (RAE).
• Operation: The maximum value is calculated along the excluded dimension for each projection, and the results are concatenated along the channel dimension.
• Efficiency: This reduces the data size per frame by 91.9% (from 260 MB to 21 MB) compared to loading the full tensor, while preserving the direct relationship between spatial bins and velocity/elevation features.

B. Backbone Architecture

The network utilizes a 2D CNN Encoder-Decoder structure (similar to U-Net) rather than computationally expensive 3D CNNs:

• Encoder-Decoder: Uses skip connections to fuse high-level semantic features with fine-grained localization.
• Attention Mechanisms: A Convolutional Block Attention Module (CBAM) is integrated into the skip connections. It applies both Channel Attention and Spatial Attention to preserve rich Doppler and Elevation features across multi-scale levels.
• Dilated Residual Neck: A neck module using dilated convolutions (with dilation rates 1, 2, 3) expands the receptive field to capture context from neighboring Range-Azimuth bins.

C. Decoupled Detection Heads

The network employs two separate heads operating on the processed feature map ($M'_{ra}$):

1. Center-Point Classification Head: Predicts a confidence map in the Range-Azimuth (polar) domain. This preserves the native coordinate system of the Radar data.
2. Bounding Box Regression Head: Predicts 8 parameters for rotated 3D bounding boxes. These are regressed in the Cartesian world frame.
   • The system converts polar center-points to Cartesian coordinates.
   • It predicts offsets ($\Delta x, \Delta y, \Delta z$) for sub-bin resolution, along with dimensions ($l, w, h$) and orientation ($\sin \theta, \cos \theta$).
   • Non-Maximum Suppression (NMS) is applied to finalize detections.

D. Loss Function

The model is trained using a composite loss:

• Focal Loss: Optimized for center-point prediction (weighted 2x).
• Gaussian-Wasserstein Distance (GWD) Loss: Treats bounding boxes as Gaussian distributions to handle rotated 3D boxes efficiently and differentiably.
• Smooth L1 Loss: Added to the GWD to prevent the model from making conservative (under-sized) predictions, ensuring full coverage of ground truth boxes.

3. Key Contributions

1. Efficient 3D Projection: A novel method to project 4D Radar tensors into 3D representations that reduce memory usage by ~92% while retaining Doppler and Elevation features.
2. Lightweight Attention Network: A custom encoder-decoder backbone using 2D CNNs and CBAM, avoiding the high parameter count of 3D CNNs while effectively exploiting low-level and high-level Radar cues.
3. Hybrid Coordinate Detection: A decoupled head design that performs classification in the native Range-Azimuth domain and regression in the Cartesian domain, optimizing for both detection accuracy and geometric precision.
4. Comprehensive Evaluation: Extensive testing on the K-Radar dataset (the only large-scale dataset with 4D RADE tensors and 3D annotations across multiple weather conditions).

4. Experimental Results

The model was evaluated on the K-Radar dataset across various weather conditions (Normal, Overcast, Fog, Rain, Sleet, Light/Heavy Snow).

• Performance vs. Baselines:
  • Outperformed the K-Radar baseline by 16.7% in overall 3D Average Precision (AP).
  • Surpassed the current state-of-the-art Radar-only model (RTNH+) by 6.5% in overall AP.
  • Achieved the best performance in Fog conditions (85.4% AP3D), significantly outperforming LiDAR-based approaches which struggle in fog.
• Adverse Weather Robustness:
  • While LiDAR and Camera-Lidar fusion models perform well in normal weather, their accuracy drops drastically in fog, rain, and snow.
  • RADE-Net maintained or improved performance in adverse weather, outperforming LiDAR approaches in almost all adverse weather categories.
• Efficiency:
  • Inference latency on an NVIDIA L40S GPU: 49ms (backbone) + 31ms (neck) + 18ms (head) = ~98ms per frame (approx. 10 FPS).
• Ablation Study: Confirmed that the combination of CBAM, the dilated neck, and the expanded heads is critical for performance. The 128-channel feature dimension was found to be sufficient, as increasing it did not yield further gains.

5. Significance

This work represents a significant step forward in Radar-only perception for autonomous driving. By demonstrating that Radar sensors can outperform LiDAR in adverse weather when paired with advanced deep learning architectures, the paper challenges the reliance on optical sensors for safety-critical applications in poor visibility.

The proposed RADE-Net offers a practical solution for embedded systems by drastically reducing memory requirements through efficient tensor projection, making high-performance 4D Radar processing feasible for real-time autonomous vehicles without the need for expensive sensor fusion. The code is open-source, fostering further research into Radar-based perception.