Enhancing Traffic Safety in Autonomous Driving Assistance Systems Through Efficient Trajectory Prediction

Abstract

Title: Enhancing Traffic Safety in Autonomous Driving Assistance Systems Through Efficient Trajectory Prediction newlineAbstract: Trajectory prediction is a critical capability for safe and reliable decision-making in autonomous vehicles operating in dynamic traffic environments. This thesis presents a structured progression of deep learning based trajectory prediction models aimed at enhancing computational efficiency, prediction accuracy, and safety assurance for autonomous driving assistance systems. newlineThe study begins with baseline implementations of SimpleRNN, GRU, and LSTM models evaluated on the NGSIM and highD datasets using Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and inference time. Although these models effectively capture temporal motion patterns, their performance degrades in scenarios requiring spatial awareness and interaction modeling among surrounding vehicles. newlineTo overcome these limitations, four novel architectures are progressively introduced. The Distributed Discriminator-Based BiLSTM (DD-BiLSTM) enhances temporal dependency learning through bidirectional memory and a distributed discriminator that improves trajectory plausibility, but suffers from higher computational cost and limited spatial representation. The Hybrid Feature-Coupled BiLSTM (HF-BiLSTM) addresses these issues by fusing spatial and temporal features using a dual-stream attention mechanism, reducing redundancy and improving precision. The Zero-Attention Spatio-Temporal ConvBiLSTM (ZST-CBTM) further refines feature selection by suppressing irrelevant spatial and channel-wise information, resulting in lower complexity and improved generalization. newlineThe final model, Fused Spatial-Temporal Attention BiLSTM (Fused STA-BiLSTM), integrates spatial and temporal attention within a unified framework to jointly model motion dynamics and contextual interactions. A post-prediction safety validation layer based on Time-to-Collision and safe-distance constraints ensures risk-aware trajectory filtering suitable for real-time deployment. newline newline newline newline newlineKeywords: Trajectory Prediction, Autonomous Vehicles, Deep Learning, BiLSTM, Attention Mechanism, Safety Validation newline newline newline