biLSCCS: modular dynamical on-road objects trajectory prediction approach

Dynamical object trajectory prediction is one of the most sophisticated tasks in computer vision, but is a highly urgent problem applied to robotics and autonomous vehicles. In this paper the novel dynamical object trajectory prediction framework biLSCCS is presented based on a six-step approach: object detector based on YOLOv5, bidirectional LSTM encoder, mean-shift multimodal clustering, 4-layer MLP-based classification, synthesis, and bidirectional LSTM decoder. The proposed approach has been tested on two benchmark datasets for object tracking and trajectory prediction, namely ETH and Stanford Drone, which involve on-road dynamic objects observed from aerial view. The experimental findings suggest that the biLSCCS approach demonstrates competitive accuracy and robustness performance in comparison to the SGAN and STAR methods. Specifically, the approach achieves ADE, FDE, and IoU scores of 0.32, 0.72, and 0.55, respectively, on the ETH dataset, and 0.22, 0.51, and 0.49, respectively, on the Stanford Drone dataset.