Volume 177

As urban traffic congestion continues to intensify, predicting short-term traffic flow has become essential to enabling real-time control in intelligent transportation systems (ITS). However, traditional models face significant limitations in capturing the spatiotemporal and nonlinear characteristics of traffic data. Long Short-Term Memory (LSTM) networks, with their gated mechanisms, can effectively model long-term dependencies and periodic patterns in traffic flow. The accuracy of these predictions directly influences decision-making in scenarios such as traffic guidance and emergency management, offering substantial practical value for improving road network efficiency. This study constructs an optimized LSTM model to evaluate its effectiveness in short-term traffic prediction and to compare predictive performance across different time granularities. A dual-layer LSTM architecture is employed, incorporating the Adam optimizer, Dropout, and early stopping as regularization strategies. Using urban traffic monitoring data from the United States, both hourly and daily prediction models are developed for experimental validation. Results indicate that the hourly prediction model (MSE = 0.0709) markedly surpasses the daily model (MSE = 0.2987), effectively identifying recurring patterns like rush periods in the morning and evening. These outcomes offer a practical solution for adaptive traffic regulation.

Show more

Semantic segmentation has undergone a remarkable transformation from traditional computer vision approaches to sophisticated deep learning architectures, culminating in the revolutionary capabilities introduced by foundation models. This comprehensive survey examines the technical progression of semantic segmentation methodologies, with particular emphasis on vision foundation models, such as the Segment Anything Model (SAM) and Contrastive Language-Image Pre-training (CLIP). This paper systematically analyzes how these large-scale pretrained models enable previously unattainable capabilities, including zero-shot learning and cross-domain generalization while identifying persistent challenges regarding computational efficiency and boundary precision. The investigation encompasses critical applications across medical imaging, remote sensing, and video understanding domains, revealing both transformative benefits and technical limitations. It concludes that foundation models represent a fundamental paradigm shift requiring hybrid approaches that effectively combine general capabilities with domain-specific optimizations.

Show more