🔢 mathematics

A Stochastic Birth-and-Death Approach for Street Furniture Geolocation in Urban Environments

This paper proposes a probabilistic framework utilizing energy maps and a stochastic birth-and-death optimization algorithm to achieve precise geolocation of street furniture in complex urban environments, validated through a realistic simulation of Dublin's street lighting infrastructure.

Original authors: Evan Murphy, Marco Viola, Vladimir A. Krylov

Published 2026-05-18

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Evan Murphy, Marco Viola, Vladimir A. Krylov

Original paper licensed under CC BY 4.0 (http://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 draw a perfect map of every streetlight, sign, and pole in a busy city like Dublin. You have a bunch of photos taken from different angles, and a computer has already tried to spot these objects in the pictures. But here's the problem: the computer isn't perfect. Sometimes it sees a streetlight where there isn't one (a false alarm), and sometimes it guesses the location is slightly off (like saying a pole is in the middle of a sidewalk when it's actually on the grass).

This paper proposes a clever new way to clean up that messy data and find the exact right spots for these street objects. Think of it as a game of "Hot and Cold" played by a very smart, slightly chaotic, but ultimately very organized team.

The "Energy Map": A Landscape of Hills and Valleys

The authors create a digital "energy map" of the city. Imagine this map as a hilly landscape:

Low valleys are good places for street furniture (like where a pole should be).
High mountains are bad places (like the middle of a building, a river, or a busy road where a pole can't exist).

They build this map using two types of clues:

The Camera Clues: When two cameras spot the same object, their "lines of sight" cross. Where they cross is a potential spot. If the cameras are confident, it's a deep valley. If they are unsure, it's a shallow dip.
The City Rules (GIS): They layer in a digital map of the city (like Google Maps). If a spot is inside a building or a park, the map turns that area into a "mountain" so the algorithm knows, "Nope, a streetlight can't go there."

The "Birth and Death" Game

Now, how do they find the best spots on this hilly map? They use a method called Stochastic Birth-and-Death.

Imagine a room full of invisible "ghost" streetlights.

Birth: Every now and then, the algorithm randomly "gives birth" to a new ghost streetlight in a random spot on the map. It's more likely to be born in a valley (a likely spot) than on a mountain peak.
Death: Then, the algorithm looks at all the ghosts. If a ghost is standing on a mountain (a bad spot) or if two ghosts are standing on top of each other (clumping), the algorithm might "kill" (remove) one of them.

This happens over and over again, thousands of times. The process is guided by a "temperature" that slowly cools down.

Hot (Early stage): The system is chaotic. It tries wild ideas, moving ghosts around freely to explore the whole map.
Cold (Late stage): The system gets picky. It only keeps ghosts that are in the deepest valleys and removes any that are in the way of others.

Eventually, the "ghosts" settle down into the most perfect, stable arrangement possible. This final arrangement is the algorithm's best guess at where the real street furniture is.

Why is this better than the old way?

The paper compares this new method to an older technique (called MRF).

The Old Way: It was like trying to solve a puzzle with rigid pieces. It was good at removing duplicates but couldn't easily use the "City Rules" (like knowing a pole can't be inside a building).
The New Way (SBD): It's flexible. Because it uses the "Energy Map," it can instantly say, "Oh, this spot looks good based on the cameras, but it's inside a building, so I'll move it."

The Results

The authors tested this in a simulation of Dublin's city center using data from 680 real streetlights. They added different levels of "noise" (errors) to the camera data to see how well the system could handle mistakes.

Accuracy: The new method found the streetlights more accurately than the old method, especially when the camera data was clean.
Stability: Even though the "birth and death" process is random, running the simulation 10 times gave very similar results, proving the method is reliable.
Handling Mistakes: When the camera data was very messy, the new method was still able to find the right spots, though it was slightly less precise than when the data was clean.

The Bottom Line

The paper claims that by mixing camera data with city maps and using this "birth and death" game, they can create a much more accurate map of urban street furniture. This helps cities know exactly where their infrastructure is, making it easier to fix broken lights or plan new ones. The code for this "game" is available for anyone to try on GitHub.

Technical Summary: A Stochastic Birth-and-Death Approach for Street Furniture Geolocation

Problem Definition
The paper addresses the challenge of precisely geolocating street furniture (e.g., road signs, telegraph poles, street lights) within complex urban environments. While deep-learning pipelines exist for detecting objects from street-level and satellite imagery, the subsequent process of accurately locating these positive detections is inherently noisy. Existing methods, such as those based on Markov Random Fields (MRF) applied to ray intersections, often lack flexibility in proposed object locations and struggle to integrate Geographic Information System (GIS) data natively. The goal is to develop a methodology that improves geolocation accuracy by integrating detection data with spatial constraints and external geospatial information.

Methodology
The proposed framework utilizes a probabilistic approach centered on an energy map and a Stochastic Birth-and-Death (SBD) optimization algorithm.

Energy Map Construction:
The core of the method is a static energy map $D$ defined over a 2D grid. Low energy values represent favorable object locations, while high values indicate unfavorable ones. The unary energy term $U(x, r)$ for a point $(x, r)$ is derived from:
- Detection Confidence: Pairwise intersections of camera-to-object rays are computed. Each intersection is assigned confidence scores ( $c_1, c_2$ ) and depth consistency scores ( $s_1, s_2$ ) based on the original detections. These are weighted and convolved with Gaussian kernels to create a confidence/depth consistency map.
- GIS Constraints: A binary map $R$ derived from Open Street Maps (OSM) penalizes locations occupied by incompatible infrastructure (e.g., buildings, water bodies).
- The final unary energy is a weighted sum of the confidence/depth map and the GIS occupancy map.
Stochastic Birth-and-Death Optimization:
The system models object configurations as a collection of discs (center $x$ , radius $r$ ) within a bounded domain. The total energy $H(g)$ of a configuration $g$ comprises:
- Unary Term: The sum of energy values from the map for each object.
- Pairwise Interaction Term: A penalty proportional to the area of overlap between objects, discouraging duplicate detections.
The optimization follows a simulated annealing schedule governed by the Gibbs distribution. The algorithm iterates through two steps:
- Birth Step: A Poisson-distributed number of new points are sampled and added to the configuration. Their locations are sampled from the domain based on the probability distribution defined by the unary energy map.
- Death Step: Points in the current configuration are sorted by energy (highest to lowest). Each point is probabilistically removed (died) based on the energy difference between the configuration with and without that point.
The process converges when the minimal energy configuration remains unchanged for a fixed number of iterations.

Key Contributions

GIS-Integrated Energy Formulation: Unlike previous MRF approaches applied to irregular intersection grids, this method encodes spatial likelihood in a native 2D map format, allowing seamless integration of external geospatial layers (GIS, road maps) to enforce physical constraints.
Stochastic Optimization: The introduction of a birth-and-death algorithm as a computationally efficient alternative to Reversible Jump Markov Chain Monte Carlo (RJMCMC) for sampling varying-dimensional object configurations.
Scalable Framework: A methodology designed to handle the noise inherent in real-world detection data while maintaining accuracy through probabilistic modeling.

Experimental Results
The approach was evaluated using a realistic simulation based on street lighting infrastructure in Dublin city center (680 ground-truth objects). The dataset simulated street-level detections with varying levels of noise (distance and bearing errors) and false positives (contamination).

Stability: The algorithm demonstrated stability across 10 simulations per noise level. Interestingly, the results showed that the SBD algorithm produced more consistent object counts and positional clusters at higher noise levels, where detections were sparser, compared to lower noise levels where tight clustering introduced more volatility.
Performance Comparison: The SBD method was compared against a standard MRF-based clustering model [Krylov et al., 2018].
- Accuracy: SBD outperformed the MRF model, particularly at lower noise levels, achieving higher precision and recall.
- Localization: Visual analysis showed SBD correctly placing objects outside of buildings (utilizing GIS data), whereas the MRF model, relying solely on ray intersections, placed objects inside structures. SBD also successfully distinguished between closely spaced ground-truth objects where MRF merged them into a single detection.
- Metrics: At a detection tolerance of 2 meters, the model performed well, with F1-scores generally exceeding those of the MRF baseline.

Significance and Claims
The paper claims that the proposed SBD formulation can match and, in specific scenarios (particularly low-noise environments and complex spatial constraints), outperform established non-stochastic MRF-based models in terms of localization accuracy. The primary significance lies in the flexible map-based representation, which enables the seamless integration of external geospatial information (GIS, road networks, placement constraints) to improve contextual awareness. This makes the approach particularly suitable for real-world applications in infrastructure monitoring and management by local authorities and private stakeholders. The authors note that while the current implementation uses simulated data to control noise variables, the framework is designed to be scalable for larger urban areas and adaptable to multimodal data sources in future work.