MS-BEV: Multi-Scales Transformer for Object Enhanced 3D Object Detection

Perceptual tasks are crucial in the field of autonomous driving and robot applications. At present, BEV perception algorithm plays a major role in multi view 3D object detection. Because of its intuitive expression and easier fusion, there are still some problems. In recent years, by adding timing fusion to make up for the limitations of single needle sensing, the accuracy and continuity of the sensing algorithm have been improved. In order to solve the problem of information loss in the process of time series fusion, this paper proposes a transform based network MS-BEV. By combining the time characteristics of time fusion and object enhancement mechanism, it enhances the understanding of dynamic object movement and solves the Long-Distance problem. Firstly, we implement a decoder layer that integrates the camera intrinsics and supports multi-scale processing, by leveraging the geometric properties of visual images. Then, we designed an object centered guided aggregation decoder to accelerate model convergence. Finally, we utilize the Gated Recurrent Unit (GRU) to achieve the feature alignment and fusion of objects. We conduct extensive experiments on the nuScenes dataset to validate the merits of our proposed MS-BEV.