In recent years, advancements in multi-camera 3D object detection technologies have made significant progress, enhancing the accuracy and efficiency of detection and tracking objects in complex and dynamic environments. A key challenge in this field is improving models’ ability to focus effectively on relevant spatial locations in the bird's eye view while filtering out irrelevant or noisy features. This research proposes a new approach to strengthen the transformer module by integrating the attention mechanism for autonomous vehicles. This approach enables the model to focus on important regions and ignore irrelevant regions, resulting in enhanced performance in multi-camera 3D object detection tasks in cluttered environments. The proposed approach consists of four blocks: an image-view encoder, a vision transformer, a BEV encoder and a special-task head. The proposed method outperforms existing image-based methods for 3D object detection in bird's eye view and 3D space, and achieves high performance in BEV detection and improved performance in 3D detection. eva‎luation of the nuScenes dataset shows consistent performance at different difficulty levels, highlighting potential challenges related to camera time delay in real-time systems. Therefore, the time delay of 0.083s of the camera led to a drop of 3.19% mAP and 8.4% NDS, and the significant effect of the time delay can be observed. When the time delay increases to 0.3s, their performance drops significantly to 26.08% mAP and 36.54% NDS.

محمدرضا احمدزاده

شيرين بقولي زاده , رسول امير فتاحي ورنوسفادراني