Research on image recognition in autonomous driving systems based on the REVA-DETR (real-time enhanced vision analysis detection transformer) algorithm

To navigate safely through complicated traffic systems, autonomous driving technologies depend on accurate object detection in real time. The algorithms currently being used to identify objects have difficulty finding the right balance between accuracies and computation efficiencies, especially when analyzing high-resolution images under varying weather and lighting conditions. To address these issues, this article presents Real-time Enhanced Vision Analysis Detection Transformer (REVA-DETR), a new approach to real-time object detection created specifically for use in autonomous driving systems. Four new technologies were integrated into the REVA-DETR framework to overcome the limitations of existing technologies: Multi-Scale Dilated Receptive Network (MS-DRNET) was developed to improve the ability to identify objects using multiple scales and obtain more context; Efficient Additive Attention (EAA) enabled for the first time the use of linear-complexity feature interactions with existing attention mechanisms; FusBoost Net was engineered to allow for multi-scale feature fusion; Minimum Point Distance IoU loss (MPD-IoU) permitted for improved bounding box regression; and lastly an experimental evaluation of REVA-DETR on both Berkeley DeepDrive 100K (BDD100K) and Self-supervised Object Detection and Annotation 10M (SODA10M) datasets provided evidence of superior performance relative to the original RT-DETR-R18 model with mean Average Precision at 50% IoU (mAP50) equal to 0.634 and mean Average Precision at 50%-95% IoU (mAP50-95) equal to 0.407 and therefore equal to improved performance of 4.96% and 24%, respectively, as well as providing continuous real-time operation at 33.2 frames per second (FPS). Additionally, the cross-dataset validation provided evidence of the ability to generalise well and balance the trade-off between detection accuracies and real-time computation needs necessary for effective deployment of autonomous driving systems.