Aiming at the difficulty of object detection and occlusion recovery in complex dynamic scenes of autonomous driving, this paper proposes a multimodal data fusion perception framework that integrates Bird's eye view (BEV) representation and three-dimensional occupancy network (OCC). The occlusion region is filled by the pseudo-point cloud generated from multi-view images, the dynamic trajectory network is predicted by introducing the time series modeling, and the multi-modal feature fusion weights are adjusted based on the scene complexity to realize the effective decoupling detection of dynamic and static objects. In the feature fusion stage, the sparse attention mechanism is used to highlight the key areas and enhance the perception of occluded objects. By constructing a dynamic detector and a static detector, the tasks of object recognition and trajectory prediction are completed respectively. The test results on NuScenes data set show that the proposed method is superior to the existing mainstream methods in target detection accuracy and occlusion recovery, and improves the robustness and multi-task perception ability of the system in complex environments, which provides an effective reference for the design of high-precision and high-reliability automatic driving perception system.