近年來由於自駕車相關技術持續蓬勃發展，其應用的交通場景亦趨於複雜，同時也考驗自駕系統中負責觀察外界環境並進行推理的感知系統所具備的泛化性及穩健性。而其中的三維多目標檢測（Multiple Object Detection )功能，更是感知系統中不可或缺的關鍵部分，其負責檢測自身車輛周遭所有物體資訊，以提升後續路徑規劃等任務決策之安全性。
為提升感知能力之準確性與穩定性，當今自駕車常配備多樣各具互補性質之感測器，以應對不同天氣條件、物體遮蔽等可能影響行駛安全之不利因素。因此，如何有效率地融合多種感測器數據依然是目前待解決的重要課題之一。
考量各項感測器部署及後續數據處理成本，本研究選擇使用相機以及毫米波雷達作為主要數據來源，並為此提出一基於交叉注意力機制的深度學習模塊，以實現兩種感測器模態特徵級別的融合方法，進而解決雷達固有的多路徑傳播噪音以及深度不準確性問題，同時提升感測器融合效率，最後應用於多目標檢測任務。期望藉由毫米波雷達所提供之徑向速度與三維狀態資訊，減少單純使用影像所造成資訊缺失，來提升物體檢測精度，並能夠在運作速度及準確度上取得平衡。

In recent years, the rapid advancement of autonomous driving technology has been circumventing increasingly complex traffic scenarios and challenging the generalization and robustness of perception systems. Among of them, 3D multiple object detection is a crucial task responsible for identifying ambient targets to provide safe-path planning and actionable decision-making.
To enhance perception accuracy and reliability, autonomous vehicles often employ complementary sensors to handle adverse conditions such as occlusion or poor-weather conditions. However, efficacy in fusing data based on multiple sensors is still the major challenge research objective.
This thesis focuses on using both cameras and millimeter-wave radar as the primary inputs to establish a deep learning module based on a cross-attention mechanism for feature-level sensor fusion. The proposed method addresses radar-specific issues like multipath noise and depth inaccuracy, improving fusion effectiveness. By leveraging radar’s radial velocity and 3D information, the approach enhances detection accuracy while maintaining a balance between speed and performance.