Abstract:
In the field of Intelligent Transportation Systems (ITS), traditional combinations of single radar and checkpoint cameras suffer from high costs and difficulties in data fusion. The radar-visual integration system, which integrates millimeter-wave radar and cameras, optimizes costs and deepens data fusion, gradually becoming a key device in the perception layer of ITS. This paper discusses the use of RFchip cascading technology, selecting the AWR2243 dual-chip solution to enhance millimeter-wave radar performance. Coupled with professional-grade security chips, a radar-visual integration system with long-distance detection capabilities is designed. On the algorithm level, a multi-sensor fusion strategy is employed, and the YOLOv8n target detection algorithm is improved by introducing the RepViT backbone feature extraction network and using the SIoU loss function to optimize bounding box regression accuracy. The improved model reduces the parameter count by 24%, shrinks the model size by 22%, and significantly improves computational efficiency while maintaining high accuracy. Through real-world traffic scenario testing, the system achieves 360° monitoring of traffic intersections, with a maximum detection range of 350 meters and a detection accuracy of over 97%, meeting the design requirements for long-range radar-visual integration systems and showing strong engineering application prospects.

1. Introduction
As urbanization accelerates and the number of vehicles increases, road traffic conditions have become more complex. Intelligent Transportation Systems (ITS), which enhance traffic management efficiency and safety, are receiving increasing attention in terms of technological evolution. Among various traffic perception sensors, the radar-visual integration system combines the advantages of millimeter-wave radar and video cameras. This fusion provides radar’s precise distance and speed measurement, which is unaffected by light or weather conditions, and the camera’s intuitive target recognition and rich information. Compared to traditional separate radar and camera deployments, the radar-visual integration system offers significant advantages in hardware cost, data synchronization, and system integration, enabling more efficient traffic state perception and event detection.

In recent years, millimeter-wave radar technology and radar-visual fusion systems have developed rapidly. Millimeter-wave radar uses the Doppler effect to detect target speed and distance, and its performance is influenced by factors such as operating frequency band, chip architecture, and antenna design. Based on detection distance, millimeter-wave radar is categorized into short-range and long-range types. Data processing often uses FPGA or FPGA+DSP architectures to balance processing speed and system flexibility. In terms of frequency bands, 24GHz and 77GHz radar are widely used. The 77GHz band is particularly suited for long-range traffic monitoring scenarios due to its shorter wavelength, smaller antenna size, greater detection range, and higher resolution.

However, existing radar-visual integration system design studies are limited, and most solutions face challenges in detection range or cost. To address these issues, this paper proposes a design solution for a radar-visual integration system with long-range detection capabilities, based on 77GHz millimeter-wave radar and RFchip cascading technology. The paper details its hardware architecture, algorithm optimization strategies, and tests the feasibility and reliability of the system.

2. System Design Overview

The radar-visual integration system is divided into two main parts: radar and video. The video section includes the main control chip and image acquisition module, using a 5MP sensor to ensure image quality. The radar section uses the AWR2243 dual-chip cascading solution to form a 6-transmit, 8-receive antenna array, with synchronization signals enabling master-slave mode cooperation. To improve RF signal quality, the system uses a dedicated power management chip for powering the RFchips. The video DSP controls the sensor to collect video data and integrates radar and video data fusion internally.

2.1 Video Section Design
The video section uses a professional-grade security chip, H13403V100, connected to the 5MP sensor via the MIPI protocol, enabling high-definition video data acquisition and transmission. This chip provides strong image processing and data scheduling capabilities, supporting multi-sensor fusion algorithms.

2.2 RFchip Module Design
The millimeter-wave radar RFchip cascading is an effective method to enhance system performance. By synchronizing multiple RFchips, transmission power and receiving sensitivity can be improved, resulting in better detection range and angular resolution. The AWR2243 is selected as the core RFchip, operating in the 76–81GHz frequency band with high integration and excellent RF performance. A single AWR2243 supports 3 transmit and 4 receive antennas, and cascading two chips expands this to 6 transmit and 8 receive, significantly improving system flexibility and detection accuracy.

2.3 Radar Data Acquisition Module Design
Radar data acquisition uses the AM2732 DSP + FPGA solution. AM2732 controls the master-slave AWR2243 chips via the SPI interface, receives LVDS data from the front-end sensors, and performs basic processing such as target distance and speed measurement. The FPGA is responsible for radar data acquisition and forwarding, transmitting data to the video control board via LVDS protocol, and completing radar-visual data fusion in the main DSP.

3. Algorithm Design

The core of the radar-visual integration system lies in the multi-sensor fusion algorithm. The system implements radar point cloud and video image fusion processing in the main DSP, including target detection, tracking, and data association.

3.1 Radar Target Detection
Radar target detection uses the DBSCAN clustering algorithm, which divides samples based on density, effectively recognizing non-spherical distributed target points, suitable for multi-target and multi-shape detection in traffic scenarios.

3.2 Image Target Detection
Image target detection is based on the YOLOv8n model, which includes Backbone, Neck, and Head components. To further enhance detection accuracy and model efficiency, this paper introduces the RepViT backbone network and uses the SIoU loss function to optimize bounding box regression.

4. Conclusion

This paper proposes a long-range radar-visual integration system design based on AWR2243 dual RFchip cascading technology. The system improves antenna performance through hardware cascading, combined with multi-sensor fusion algorithms and an improved YOLOv8n detection model. The system achieves detection accuracy over 97%, with the maximum detection range reaching 350 meters, meeting the requirements of long-distance traffic monitoring in ITS scenarios. The test results validate the system’s stability and reliability, offering great engineering application value.