Deep Learning Frameworks for Real-Time Passenger Detection and Tracking

This research develops and evaluates two deep learning frameworks—DynaTrackNet and MemoFuseNet—for detecting and tracking passengers in crowded visual environments. A complete web-based platform is constructed using HTML, CSS, JavaScript for the front-end, and Python with the Flask framework for the back-end. SQLite serves as the database for user management. The system includes five functional modules: Home, Register, Login, Dashboard, Live Detection, and Logout. DynaTrackNet focuses on motion dynamics by predicting target positions across sequential frames, ensuring continuous tracking even under partial obstructions. MemoFuseNet employs a memory-augmented structure that stores long-term appearance features, enabling re-identification of individuals after temporary disappearance. The Crowd Detection dataset from Roboflow provides annotated images labeled with a single class Crowd to train and validate both algorithms. The platform allows users to switch between the two detection engines via an interactive dashboard and observe live detection outputs with bounding boxes and trajectory lines. Performance metrics include tracking consistency, identity switch frequency, and processing latency. Results indicate that DynaTrackNet offers smoother motion continuity, while MemoFuseNet reduces identity mismatches in dense scenarios. The system provides a controlled environment for comparative analysis of motion-based versus memory-based tracking architectures. This work contributes a modular, accessible tool for research in passenger tracking using deep learning.

Keywords: Passenger detection, object tracking, DynaTrackNet, MemoFuseNet, deep learning, Flask framework, SQLite, crowd dataset, motion prediction, memory-augmented networks