Future instance prediction in hybrid spatial-Temporal birds-eye-view representation for vision-centric autonomous driving

Understanding surrounding dynamics and predicting the future states of instance-level targets are crucial for ensuring safe navigation in autonomous driving. The traditional modular paradigm, which cascades multiple discrete modules through rule-based interfaces, is challenged by information loss and compounded errors. End-To-end motion prediction can directly access sequential sensory data, bypassing redundant intermediate components by utilizing a unified network. Given the advantages of low-cost resources and rich semantics provided by cameras, this work focuses on investigating vision-centric joint perception and prediction for instances of interest. Specifically, we proposed a hybrid spatial-Temporal pyramid module based on 2D convolution kernels and attention mechanisms within the Bird's-Eye View (BEV) plane, which hierarchically extracts global and local contextual information. Additionally, a feature-level motion excitation module is designed to capture and augment the representation of motion-Aware channels. Experiments on the large-scale nuScenes dataset demonstrated that the proposed method outperforms existing algorithms and more accurately predicts target movements over the next 2 seconds across two scene ranges with different resolutions.