Pedestrian Trajectory Prediction and Pose Estimation Considering Behavioral Characteristic Relationships and Noise Inhibition

In the domain of pedestrian trajectory prediction(PTP) from a roadbed perspective, the visibility of pedestrian feature points is inevitably compromised by external noise interference, impacting both pedestrian pose estimation(PPE) and PTP. This paper presents an innovative model for PTP. The model not only tackles noise interference but also takes into account the inherent correlation between pedestrian pose features and trajectory coordinates. To tackle the challenge of noise interference during pedestrian crossing, we reframe it as an anomalous feature detection problem using the Graph Deviation Network (GDN). Subsequently, we enhance the Long Short-Term Memory (LSTM) module by incorporating a time-domain anomaly suppression module, resulting in the development of an Anomaly Inhibition-LSTM (AI-LSTM) with robust noise suppression capabilities. Finally, by integrating the predicted values of behavioral pose and trajectory position, considering the behavioral characteristic relationship resolved by the GDN algorithm, we achieve accurate prediction and pose estimation of pedestrian crossing trajectories amidst noise interference. Experimental results demonstrate superior performance of our algorithm in the PPE task when compared to GDN and LSTM algorithms. In the PTP task, our algorithm exhibits performance comparable to the Transformer-based method, with the added advantage of improved interpretability.