💻 computer science

Real-Time Urban Vehicle Tracking via Wireless Sensor Networks Using Kalman Filter and Random Forest Prediction

This paper proposes a real-time urban vehicle tracking framework that integrates YOLOv8 detection, an adaptive Kalman Filter, and a Random Forest predictor within a wireless sensor network to achieve high-accuracy, identity-stable tracking under dynamic traffic conditions with low latency.

Original authors: T Vairam, S. Venkatesan

Published 2026-08-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: T Vairam, S. Venkatesan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to keep score in a chaotic, high-speed game of tag played in a crowded city square. The players are cars, the field is a busy intersection, and the "taggers" are cameras watching from above. In the world of computer science, this is called Multi-Object Tracking. The goal is simple: watch a video feed, spot every car, and follow each one individually without losing track of who is who, even when they zoom past each other, hide behind trucks, or suddenly swerve.

To do this, computers usually rely on two main tools. First, they need to see the cars, which is like a referee spotting players in a crowd. Second, they need to predict where those cars will go next, even if the camera can't see them for a split second. This prediction part is tricky because cars don't always move in straight lines; they brake, turn, and speed up unpredictably. If the computer guesses wrong, it might think two different cars are actually the same one, or it might lose a car entirely when it gets blocked by a bus. Getting this right is crucial for smart cities, traffic management, and keeping roads safe, but it's a massive challenge when the video feed is shaky, the lighting is bad, or the network connection is spotty.

This paper introduces a new way to play that game of tag, designed specifically for a network of wireless cameras (called Wireless Sensor Networks) that might not have super-fast internet connections. The authors, T Vairam and S. Venkatesan, propose a "hybrid" system that acts like a super-smart referee team. Instead of relying on just one method, they combine a classic, math-heavy prediction tool called a Kalman Filter with a machine learning brain called a Random Forest.

Think of the Kalman Filter as a veteran coach who knows the rules of physics perfectly. It assumes cars move smoothly and predicts their next position based on their current speed. However, this coach gets confused when a car suddenly slams on the brakes or makes a sharp turn. To fix this, the team adds a "scout" trained on a Random Forest model. This scout has studied thousands of hours of past traffic videos and learned the messy, real-world habits of drivers—like how they actually behave when they are in a hurry or stuck in a jam. The scout doesn't just guess; it corrects the coach's predictions based on what it has learned from history.

The paper suggests that by blending these two approaches, the system becomes much better at handling the chaos of city traffic. The "coach" (Kalman Filter) handles the smooth, steady driving, while the "scout" (Random Forest) steps in to fix the predictions when things get weird or sudden. To make this even smarter, the system also has a special trick: it constantly adjusts its own "confidence meter" (called the Q-matrix). If the cars are moving erratically, the system knows to trust the scout more; if they are cruising steadily, it trusts the coach. This allows the system to stay accurate even when the video feed is noisy or when cars are hidden from view for a few frames.

The researchers tested this idea using real traffic videos and simulated city environments. They found that their hybrid team performed impressively. In their tests, the system achieved a high overall tracking score (MOTA) of 0.91, which measures the system's ability to correctly detect vehicles, avoid false alarms, and maintain their identities without mixing them up. It also kept the identity of each car consistent, only swapping the ID of a vehicle about 11 times in a long sequence (called ID switches). This is better than using just the "coach" alone or just the "scout" alone.

Perhaps most importantly for real-world use, the system is fast and lightweight. It runs on modest computer chips found in edge devices, processing video at about 13.9 frames per second with a delay of only 71 milliseconds per frame. This means it could actually work on a network of wireless cameras in a real city, even if the internet connection is a bit shaky or if some data packets get lost along the way. The authors suggest that this approach offers a practical, efficient way to keep an eye on urban traffic without needing massive, expensive supercomputers, making it a strong candidate for future smart city infrastructure.

Technical Summary: Real-Time Urban Vehicle Tracking via Wireless Sensor Networks Using Kalman Filter and Random Forest Prediction

Problem Statement

Urban traffic surveillance faces significant challenges in real-time multi-vehicle tracking, particularly within the constraints of Wireless Sensor Networks (WSNs). Key difficulties include non-linear vehicle motion, sudden trajectory changes, frequent occlusions, and measurement noise inherent in CCTV feeds. Traditional tracking methods often struggle with these dynamic conditions:

Conventional Kalman Filters (KF): While computationally efficient, they rely on fixed process noise covariance (Q-matrix) and linear motion models, leading to poor performance during abrupt maneuvers or in noisy environments.
Deep Learning Trackers: Models like DeepSORT offer high accuracy through appearance features but are often computationally expensive, introducing latency that hinders real-time deployment on edge devices or resource-constrained WSNs.
Hybrid Limitations: Existing hybrid approaches often lack dynamic adaptivity, failing to couple classical filters with machine learning models in a confidence-driven manner that accounts for real-time system latency and packet loss.

The paper identifies a gap for a lightweight, real-time system that combines adaptive filtering with explainable machine learning to maintain identity stability and tracking accuracy in unpredictable urban traffic without the heavy computational overhead of deep recurrent networks.

Methodology

The proposed framework integrates three core components: YOLOv8 for detection, an Adaptive Kalman Filter for state estimation, and a Random Forest (RF) regressor for motion prediction correction.

1. Object Detection

The system utilizes YOLOv8 (nano version) for real-time vehicle detection. This model was selected for its balance of inference speed and accuracy in dense traffic scenes. It processes video frames to output bounding boxes, class labels, and confidence scores. Only detections exceeding a specific confidence threshold are passed to the tracking module to minimize noise propagation.

2. Adaptive Kalman Filter (KF) with Dynamic Q-Matrix

The core tracking engine employs a Kalman Filter with a state vector $X = [x, y, v_x, v_y]$ . To address the limitations of static noise modeling:

Adaptive Q-Matrix: The process noise covariance matrix ( $Q$ ) is dynamically updated at every frame based on the observed variance in vehicle velocities. This allows the filter to increase its sensitivity to sudden accelerations or lane changes while maintaining stability during steady-state motion.
Measurement Noise (R-Matrix): The measurement noise covariance is adjusted based on the confidence score of the YOLOv8 detection, weighting high-confidence detections more heavily.
Data Association: A two-stage gating mechanism combines Intersection over Union (IoU) and Mahalanobis distance to associate detections with existing tracks, reducing false associations in crowded scenes.

3. Random Forest Motion Prediction

To correct the linear assumptions of the KF, a Random Forest regressor is trained on historical trajectory data.

Input: Features include past spatial coordinates, frame-wise velocity changes, and inter-frame displacements.
Function: The RF model learns non-linear motion patterns (e.g., sharp turns, stop-start behavior) that the KF cannot predict. It outputs a correction vector for the predicted motion.
Fusion: The final state estimate is a weighted blend of the KF prediction and the RF correction. The weighting coefficient ( $\alpha_t$ $α_{t}$ ) is context-aware, dynamically calculated based on:
1. Temporal variance of vehicle velocity.
2. Consistency of recent detection confidence.
3. The magnitude of residuals (errors) from the KF update.

4. System Integration and WSN Optimization

The system is designed for edge deployment (e.g., NVIDIA Jetson Nano) and WSN environments. It includes mechanisms for:

Occlusion Handling: Tracks are maintained via a "missed counter" during temporary occlusions (up to 5 frames) rather than being immediately terminated, reducing ID switches.
Latency Management: The architecture tolerates packet loss and network jitter by using KF state extrapolation to bridge gaps in data transmission.

Key Contributions

Hybrid Fusion Architecture: The paper proposes a novel residual-feedback fusion mechanism where the Random Forest model learns to correct the specific prediction drifts of the Kalman Filter, rather than operating as a parallel pipeline.
Adaptive Noise Modeling: The introduction of a dynamic Q-matrix tuning strategy based on real-time velocity variance and residual errors, significantly improving responsiveness to non-linear urban traffic dynamics.
WSN-Ready Design: The system is explicitly optimized for low-latency, resource-constrained environments, demonstrating robustness against packet loss (up to 25%) and network delays through predictive continuity.
Explainable Efficiency: By utilizing Random Forests instead of deep recurrent networks (LSTM/GRU), the system achieves high accuracy with lower computational complexity and better interpretability, suitable for edge devices.

Experimental Results

The system was evaluated using standard Multi-Object Tracking (MOT) metrics (MOTA, MOTP, IDF1, ID Switches) on the MOT20, KITTI, and Urban Tracker datasets, alongside a custom CCTV dataset.

Performance Metrics: The proposed Hybrid (Blended) model achieved:
- MOTA: 0.91 (vs. 0.82 for KF-only and 0.89 for DeepSORT).
- IDF1: 0.88 (vs. 0.74 for KF-only).
- ID Switches: 11 (significantly lower than KF-only's 35 and comparable to DeepSORT's 13).
- MAE/RMSE: Lowest error rates among tested configurations (MAE: 0.84 pixels).
Real-Time Capability:
- FPS: Achieved 13.9 FPS on an embedded Jetson Nano and 23.7 FPS on a desktop, meeting real-time thresholds (>10 FPS).
- Latency: Average end-to-end latency was 71 ms/frame on the embedded platform.
WSN Resilience: Simulations showed the system maintained ~10.5 FPS even under 25% packet loss in a 4G mesh topology, thanks to KF-based trajectory extrapolation.
Ablation Study: Results confirmed that both the Adaptive Q-matrix and the Random Forest correction were critical; removing either component resulted in a measurable drop in MOTA and an increase in ID switches.

Significance and Claims

The paper claims that this framework successfully bridges the gap between model-based filtering and data-driven learning, offering a solution that is both computationally efficient and highly accurate.

Practical Applicability: The system is presented as a viable solution for Intelligent Transportation Systems (ITS) and smart city infrastructure, specifically where bandwidth is limited and edge computing is required.
Robustness: The adaptive nature of the Q-matrix and the data-driven correction of the RF model allow the system to handle the unpredictability of urban traffic (occlusions, sudden stops, lane changes) better than static filters or heavy deep learning models.
Scalability: The modular design supports distributed processing across sensor nodes, making it feasible for large-scale deployment in wireless sensor networks without overwhelming central servers.

The authors conclude that while the system is robust, it relies on linear assumptions within the KF and lacks appearance-based re-identification for long-term occlusions. Future work is suggested to incorporate lightweight temporal modules (e.g., GRUs) and transformer-based embeddings to further enhance long-sequence prediction and identity preservation.