A Deep Learning-Based Vision-Sharing System with Image Stitching for Blind Spot Reduction in Vehicle-Following Scenarios

This study presents a small-scale application-oriented vehicle vision-sharing prototype for blocked-view observation in vehicle-following scenarios. The system was implemented using two DuckieBot vehicles and integrates socket-based image transmission, JPEG image compression, You Only Look Once (YOLO)-based object detection, road-plane-aligned image fusion, and fuzzy proportional–integral–derivative (fuzzy-PID) motor control. These modules are adopted as existing techniques and integrated for prototype-level experimental evaluation rather than proposed as new perception, compression, fusion, or control algorithms. Experiments were conducted under controlled small-scale indoor conditions. JPEG compression was quantitatively evaluated using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), encoded file size, and processing time. Q75 provided a mean PSNR of 39.8777 dB, a mean SSIM of 0.970522, and an average encoded size of 27.78 KB, representing a practical trade-off between reconstructed image quality and encoded data size. The YOLOv8n obstacle detector achieved a precision of 0.9724, a recall of 0.9571, an [email protected] of 0.9851, and an [email protected]:0.95 of 0.8585 on an independent test set. Image-fusion evaluation showed that α = 0.60 produced the highest global mean PSNR, whereas α = 0.90 produced the highest global mean SSIM, indicating that the preferred blending coefficient depends on the selected image-quality criterion. A system-level ablation further distinguished shared-view visualization from a warning-only configuration, with the expected obstacle information presented in all 35 positive trials and no false alarms observed in 10 negative trials. The vehicle-following experiment verified the functional operation of the complete perception-to-control pipeline. The results should be interpreted within the controlled miniature-vehicle setting and should not be directly generalized to full-scale vehicles or real-road advanced driver assistance systems.