A CVAE Combined With Diffusion Mechanism to Pedestrian Trajectory Prediction

Pedestrian trajectory prediction has always been a key issue in engineering fields such as autonomous driving. Diffusion model can be used to solve the problem of multimodal pedestrian trajectory generation, but the excessive denoising steps in the model result in slow inference speed for this type of method. To address these issues, we propose a multimodal trajectory generation model, which is a CVAE framework combined with diffusion mechanisms. This model utilizes the denoising process in the diffusion model to optimize the trajectory output by CVAE, thereby achieving higher accuracy. Meanwhile, the initial trajectory output by CVAE can significantly reduce the denoising steps in the model, thereby reducing inference time. Conducted experiments on ETH and UCY datasets, and ADE and FDE results showed that our method achieved better performance compared to previous methods. Meanwhile, our method also demonstrated better test results on the NBA dataset.