PTIN: Enriching Pedestrian Safety with an LSTM-GRU-Transformer Based Trajectory Imputation System for Autonomous Vehicles

Human Trajectory Imputation has been used for decades in surveillance, sports analytics, delivery and logistics, traffic management, etc. Recently, autonomous vehicles(AV) have entered into this realm. These vehicles have come to the transportation scenario for their safety, efficiency and convenience. These vehicles, filled with advanced sensors; analyze and navigate critical environments. GPS-based trajectory data, among these sensors, plays a vital role in motion planning. However, challenges such as signal loss due to obstructions, electromagnetic noise and technical constraints create in missing or erroneous trajectory points. Drones, positioned higher in the sky, offer advantages in signal reception and obstacle avoidance. That is why we are using inD dataset. The techniques addressed here tend to ensure reliable, real-time and accurate autonomous vehicle operation. With the stringent real-time demands of autonomous driving, efficient trajectory imputation methods are crucial. Advanced architectures including LSTM, GRU and Transformer capture temporal dependencies, intricate patterns and contextual relationships within trajectory data. Their integration leads to a novel hybrid model, coined as PTIN(Pedestrian Trajectory Imputation Network), addressing pedestrian trajectory for Autonomous Vehicle challenges. The findings contribute insights and recommendations for effective trajectory imputation, which enhance the reliability and safety of autonomous vehicle systems, facilitating their widespread adoption.

I. Introduction

Autonomous vehicles represent a transformative advancement in the realm of transportation, promising safer, more efficient, and convenient mobility solutions. Autonomous vehicles utilize advanced sensors and computing systems to navigate complex environments and make decisions without human intervention. GPS-based trajectory data is crucial for precise localization, mapping, and path planning. However, GPS-based trajectory data faces many problems, such as signal loss, interference, and technical deficiency. We are aiming for drone datasets like inD(Intersection Drone) [1]. We selected the InD dataset because every record in it was taken directly from a high-resolution drone video. Compared to sensors at ground level, the perfect drone viewing angle allows for the measurement of a full junction situation with substantially less obstruction or interference with the signals. Filling missing or erroneous trajectory data is paramount to ensuring the reliability and accuracy of autonomous vehicle systems. This research paper delves into the nuances, methodologies, and advancements in Trajectory interpolation tailored specifically for the needs of autonomous vehicles.