With the increasing prominence of road traffic safety issues, the technical demand for large truck blind spot monitoring has urgently risen. This paper proposes an innovative blind spot monitoring method that combines a lightweight deep learning model with multi-image sensor data fusion technology, conducting in-depth research and practical application for large truck blind spot monitoring. By introducing an improved YOLOv5 algorithm, modifying the loss function to SIoU, and integrating Ghostnet modules, BiFPN, and CA attention mechanisms, and by combining the Bdd100k dataset with a self-built dataset for model training and testing, the proposed improved YOLOv5 algorithm reduced model size by 15.9%, increased detection accuracy by 0.6% compared to the original YOLOv5 model, achieving an average precision of 77%, and improved inference speed by 4 FPS. This validates the effectiveness of the improved algorithm, as it slightly enhanced detection accuracy and inference speed while maintaining model lightweightness.

To address camera visual distortion issues, this paper analyzed image sensor calibration methods to achieve camera undistortion correction, obtaining camera intrinsic parameters and distortion coefficients, thereby improving image quality. To tackle the problem of large blind spot fields of view in large trucks, this paper utilized the ORB algorithm for splicing and fusing multi-image sensor video streams, expanding the blind spot detection range and enhancing the system's adaptability and robustness to complex traffic environments. Finally, an experimental platform was built to acquire camera video data, and the improved YOLOv5s algorithm model was deployed on an RK3588 development board, providing an effective technical solution for large truck blind spot detection.

To fulfill the object detection task for large vehicle blind spots, this chapter first built an experimental platform and used the V4L2 framework for camera data acquisition. Then, the improved YOLOv5 object detection algorithm model was converted into a model format supported by the NPU (Neural Processing Unit) integrated in the RK3588, and finally deployed on the RK3588 development board for inference and detection, achieving mobile-end embedded deployment of blind spot object detection.

5.1 Experimental Platform Construction The development board deployment end of the experimental platform in this paper uses the Xinmai Technology RK3588 development board, which runs Ubuntu-20.04.5. The RK3588 is a domestic processor architecture from Rockchip, equipped with an NPU (neural network processing unit), which accelerates neural network processing and is widely used in deep learning model deployment.

The debugging platform for the development board is a laptop running Windows 11 with a virtual machine running Ubuntu 20.04.6. Table 5.1 lists the hardware models and software versions of the experimental platform. Figure 5.1 shows the USB communication interface camera of the CMOS model used in the experiment.