Automotive Radar Advancements with AI – Revolutionizing Safety & Perception (Part 1)
September 5, 2023
Introduction
The Automotive Industry has witnessed significant advancements in Autonomous Driving capabilities, with Automotive Radar playing a critical role. By integrating Artificial Intelligence (AI) techniques, radar systems have experienced substantial improvements in perception, accuracy, and safety.
Key roles of AI in Automotive Radar:
- Intelligent Object Detection and Classification
- Signal Processing and Feature Extraction
- Real-time Adaptation and Decision-making
- Object Tracking and Trajectory Prediction
The traditional method of Radar Signal Processing includes acquiring the Analog to Digital Converting (ADC) Data from the Radar front end, computing 1D FFT (Fast Fourier Transform) to get range followed by Doppler (to get velocity), Azimuth (to get the angle in the horizontal direction) processing, and Elevation (to get vertical direction) followed by extraction of the point cloud. The pipeline depicted in Fig 1 is used in most Automotive Radar applications.
Fig 1. Radar data processing pipeline
What if the classification and feature extraction of the objects using AI can be done at the ADC data level which will help to improve the computation time? In this blog, we will explore how AI is transforming Automotive Radar, enhancing its performance and enabling safer Autonomous Driving.
AI Advancement in Radar Signal Processing
Radar technology that uses radio waves to detect and locate objects has immensely benefited from the application of Artificial Intelligence and Machine Learning techniques. Here are some of the use cases where AI has made significant progress in Radar Signal processing:
- Enhanced Target Tracking
- Particle Filters are applied for non-linear and Gaussian tracking scenarios.
- Tracking algorithms based on deep-learning utilize RNNs (Recurrent Neural Networks) and LSTM (Long Short-Term Memory) networks.
- Adaptive Beamforming
- Reinforcement learning is used to adopt beam patterns in real time.
- Leveraging deep learning for beamforming to reject interferences, we apply a CNN-based approach that relies on the DOA (Direction of Arrival) of the signals. It learns to transmit the next signal set by validating the received signal from a specific direction.
- Anomaly Detection
- One-class SVMs are employed to identify irregularities or anomalies within radar data. An advantage is that it lacks target labels during the model training phase.
- Autoencoders are used for unsupervised anomaly detection.
- Detect Ghost Images
- Cognitive Radar Antenna selection using Deep Learning
- The resolution of a radar is determined by the number of virtual Antennas in a radar.
- A higher number of antennas will increase the cost, size, and power consumption. This problem can be solved by using a DNN to send out an adaptive phase-modulated waveform based on the environment using a small number of virtual Antennas which will improve the resolution.
We will now look into Radar Signal Classification, one of the major improvements using AI which helps to resolve more than one problem at an early stage of Radar Data processing.
Multi-Task Learning for Radar Signal Characterization
Traditionally, classification of RF waveforms is achieved using likelihood-based and feature-based methods which require prior knowledge about the signal characteristics and thus, cannot effectively cater to new signals. Since the Neural Network-based approach does not require any prior knowledge, it can do the characterization with minimal compute power once trained.
A deep Neural Network (Multi-Task Learner) for Radio Signal Recognition is used to estimate the Pulse Width (TPW), Pulse Repetition Interval (TPRI) of a radar transmission, number of pulses (NP), and Pulse time delay (TD) which helps in the classification of the radar signal.
Data Flow and Architecture:
The Multi-Task Learner (MTL) framework is used for tackling Radar Signal Characterization (RSC) as a joint optimization problem. The proposed IQST among other architectures performs simultaneous optimization of classification and regression tasks. This will help in feature extraction from the IQ data particularly at low SNRs (Signal-to-Noise Ratios).
Fig 2. The proposed MTL architecture for RSC (Radar Signal Characterization) with IQST backbone followed by task-specific classification and regression heads.
(Source: https://arxiv.org/pdf/2306.13105v1.pdf )
- Raw IQ data with no additional preprocessing and feature transforms is passed as the input.
- Introduced IQST (IQ Signal Transformer), as shown in Fig 2, which comprises patch embedding technique to generate a sequence of 1D patch embeddings from 2×512 tensor constructed from raw IQ sequence.
- The dual-channel IQ data is flattened and then the dense linear projection is applied.
- Then it is passed to Transformer Encoder which adopts the GELU activation function and implements 3 multi-head attention blocks and 3 encoder layers.
- The output is fed as a feature map into each task-specific head to complete the MTL (Multitask Learner) model.
- The Multitask segment includes a single convolutional layer with a kernel size of 3×3 followed by a dense layer.
- Dropout rates of 0.25 and 0.5 are applied to the above layers.
- ReLU activation function is used in each head, while batch normalization is applied before the activation function.
- For the final head, a SoftMax is applied for the classification.
The modularity of the MTL design provides opportunities to add additional classification and regression tasks in future works which will help to reduce blocks in the traditional signal processing chain.
Conclusion
The integration of AI techniques has significantly advanced Automotive Radar technology, revolutionizing perception capabilities and enhancing safety in autonomous driving. With intelligent object detection and classification, improved resolution and range estimation, real-time adaptive beamforming, object tracking, trajectory prediction, environmental modeling, and rapid adaptation to changing conditions, AI continues to evolve. We can expect further enhancements in Automotive Radar, leading to safer and more efficient autonomous driving experiences.
MulticoreWare’s strengths in Automotive Radar
- Extensive knowledge in processing the ADC data into a point cloud both in x86 as well as in embedded platforms using DSP, HWA, and ARM cores.
- Ability to build a control and Data path pipeline from scratch in any platform, especially in TI soc.
- Pre-processing and post-processing of radar data which can be entered into Machine Learning models and optimized for edge devices.
- Experience in Multiple Object Tracker both in x86 as well as Embedded platforms.
For more information, please contact us at info@multicorewareinc.com
