Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception

arXiv:2611.08441 · cs.CV, eess.IV · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception".

Jane: Here's the gist of the paper. Cyclist fatality rates remain high despite policy support for active travel, and autonomous vehicles add complexity to mixed traffic The EBM regressor achieves 72.2% binary accuracy and MAE 1.011 on a 1–9 scale This system presents an edge-assisted crowdsensing pipeline that turns handlebar video into a structured sketch of each scene at the network edge, fuses it with OSM road infrastructure and rider kinematics, and scores perceived safety with an explainable boosting machine System Overview The paper introduces an edge-assisted crowdsensing system designed to monitor cyclist safety perception along a route by processing handlebar camera video at the network edge.

Tom: This pipeline involves three main stages: capture on the bicycle, attribute extraction at the network edge, and scoring wherever the attribute vector is consumed.

Jane: The core innovation is using a single vision–language model (VLM) prompt to turn each frame into structured traffic attributes.

Tom: Attribute Extraction and Fusion The attribute extraction stage utilizes a VLM to convert frames into structured traffic attributes through a single prompt, replacing specialized multi-model detection pipelines.

Title and authors: Tom: So we're looking at this paper now titled "Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception" and the authors are from Queen Mary University of London, Imperial College London, and Cambridge. Jane, what’s the big picture here for someone listening right now?

Jane: Well, Tom it’s about tackling cyclist fatality rates which stay high even with all these policies pushing for active travel and autonomous vehicles adding complexity to traffic. This system they propose is an edge-assisted crowdsensing pipeline that uses handlebar camera video to turn each frame into structured traffic attributes using just a single vision language model prompt and then scores the perceived safety with an explainable boosting machine.

Lu: What I find interesting is how they’ve moved the attribute extraction part to the network edge, which is a big deal because it means you don't have to send massive amounts of video data all the way to a central cloud for processing. This pipeline involves capture on the bicycle, attribute extraction at the edge, and then scoring wherever that attribute vector is used.

Meng: From my side I’m thinking about deployment cost and what actually breaks when you try to scale this thing up in a real city environment. The paper mentions that the fastest selfhosted model needs two point four nine seconds per frame, which is slow for continuous monitoring on a bike.

Lalam: I can look at the table they show, and it highlights how different vision language models perform on extracting these attributes through just one prompt. They found that Grok two vision reaches sixty seven point one percent accuracy on high traffic density, and that’s compared to the baseline of always predicting the majority class.

Tom: That's a solid number from them, Lalam. But what's really impressive is how they use those extracted attributes—like cycling lane presence or construction site presence—and fuse them with OpenStreetMap road infrastructure and on-bike kinematics like speed three dimensional and acceleration three dimensional to get a final safety score.

Jane: Right, so the core innovation is using that fusion stage with an explainable boosting machine, or EBM, to predict that crowdsourced one to nine safety score. And what makes it really useful for planners is that each term in the EBM equation plots as a shape function showing how a feature raises or lowers the predicted score.

Lu: That direct attribution is powerful because it means we can price infrastructure changes in rating points. For instance, they say physical lane separation is worth point zero nine seven rating points, and construction site presence costs point zero six five points even when there's physical separation present.

Meng: Pricing things by perceived safety sounds useful for cost-benefit analysis, but what about the actual operational load of running this thing at scale? I’m curious about the compute requirements for scoring once you have that attribute vector.

Tom: The scoring stage is pretty light on resources; they say the explainable boosting machine regressor runs in point zero four milliseconds on a cpu. That's extremely fast, which is good for real-time feedback as the rider moves along their route.

Title and authors: Jane: And they showed that the system reaches seventy two point two percent binary accuracy and a mean absolute error of one point zero one one on a scale from one to nine when they tested it on twelve hundred thirty six annotated frames from two London routes. That shows it’s quite accurate for what it’s trying to do.

Lu: They also demonstrated robustness, meaning no model pair separates under a tighter protocol when they use spatial blocking, and they found that the unsafe class recall is point zero eight one five while the safe class recall is point zero six two three, which tells us the model errs on the side of caution.

Meng: That’s interesting about it being biased towards caution. But what about making this continuous? I’m thinking about how you actually deploy this to serve a whole fleet of riders without draining batteries or requiring constant high bandwidth.

Tom: They addressed that by showing how re-encoding the frame to a twenty second sampling interval serves one thousand riders at fifty four megabit per second, and it costs only one point zero four accuracy points for that continuous monitoring.

Jane: That seems practical because it significantly reduces the payload by twenty three times when they re-encode the frame to a one thousand two forty pixel long edge at jpeg quality eighty five, bringing the fleet requirement down to fifty four megabit per second.

Lu: It’s a clever way to reduce bandwidth, but I do have some reservations about that continuous monitoring part. The paper itself flags that serving large vision language models at the network edge is an active research area, and they note that in which coverage participant burden and uplink cost are design variables nine, ten <ref:2611.08441#pg1>.

Meng: So, what’s the cost of running those larger models when you’re talking about deployment? Is it feasible for smaller edge devices or does it still need a serious GPU setup somewhere?

Lalam: The paper shows that the fastest selfhosted model needs two point four nine seconds per frame if you want to run it locally at the edge, which suggests that while there are faster configurations, you’re still looking at a significant processing time for real-time edge deployment.

Tom: It seems like they’ve balanced the speed of extraction on the edge against the need for a reasonable accuracy trade-off when they discuss these performance numbers. But what about those attributes themselves? Are they reliable enough to feed into this scoring system?

Jane: The reliability comes from how they trained that single vision language model prompt to turn each frame into those fourteen structured traffic attributes, and it seems pretty effective at capturing the necessary scene-level cues. They show that these five input models—cycling lane, lane separation, construction site presence, immediate crash risk from traffic, and high traffic density—are what feed the final EBM.

Title and authors: Lu: I’d be interested in seeing how those input attributes change if we were to use a different set of features instead of just those five; does swapping them out fundamentally alter the safety score predictions? That’s where you really test if the idea is robust beyond this specific setup.

Meng: If we swap out features, we immediately run into the deployment question again—does that new feature require a model that takes three seconds to process instead of less? I need something that fits on a mobile edge device.

Tom: The system’s strength seems to be in the explainability provided by the EBM, which gives us those per-feature contributions, but we have to keep an eye on how quickly those attributes can be extracted at the edge without spiking processing time too much.

Jane: It really boils down to using a single prompt for extraction instead of needing specialized multi-model detection pipelines, which is a simplification that makes deployment easier. And they show that this works even when we compare different vision language models against each other in their table.

Lu: That comparison is useful because it shows that the performance gap between the models isn't huge for the attributes they are measuring, but it’s still important to know which one you’re using for your specific application. It points toward a need for more consistent attribute quality across different edge hardware.

Meng: Consistency is key when we ship this thing out and it breaks on some kind of weird traffic pattern that the training data didn't cover. I need to know what happens when the input video quality degrades, say, because of rain or poor lighting on a bike camera.

Tom: That’s a limitation they acknowledge implicitly by focusing on these specific features like construction sites and lane separation, but they don't really detail the error propagation under severe weather conditions in this paper. They focus more on the accuracy achieved across those twelve hundred thirty six frames.

Jane: So, to wrap up for this part of the discussion, we’ve seen how they turned video into structured attributes using a single vision language model and then used an explainable boosting machine to score safety by combining those attributes with road data and rider kinematics. This gives planners a way to weigh interventions directly based on perceived risk points.

Lu: It’s a useful tool for planners because it moves beyond just accident records, which only show outcomes, and gives them subjective experience scaled into something they can use in cost-benefit analysis.

Meng: For production, the main hurdle seems to be balancing the need for high-quality attributes with the strict latency requirements of edge deployment on consumer hardware. We need to see if that point zero four millisecond scoring time holds up when we push it out to thousands of devices simultaneously.

Tom: The paper "Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception" shows a solid proof of concept for how crowdsourcing can provide actionable, explainable data on cyclist safety perception along routes. This moves the conversation from just counting crashes to understanding what people actually feel about their environment while riding.

The paper's summary: Tom: So we’re looking at the summary of this paper now, "Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception," and it boils down to taking video from a bike, processing it right there on the network edge, and using that information to give a safety score.

Jane: Exactly. The main idea is that instead of sending all those raw videos to some giant central computer in the cloud for analysis, they’re doing the heavy lifting at the location—the edge—where you want the data immediately.

Lu: It turns handlebar video into structured attributes using just one vision language model prompt instead of needing a whole suite of specialized detection models. That simplifies things a lot for deployment.

Meng: I’m thinking about that "single prompt" thing because it sounds like a huge win for fitting these kinds of models onto smaller edge devices, which is what we actually need in the field.

Tom: Right, and then they take those extracted attributes—things like whether there's a cycling lane or if there's construction nearby—and they fuse them with real map data from OpenStreetMap and how fast the rider is moving.

Jane: That fusion gives them a final safety score, which they calculate using an explainable boosting machine, or EBM. And the EBM is what makes it useful because it shows you exactly *why* it gave that score.

Lu: The EBM lets you see how each feature—like lane separation or construction site presence—contributes to the final score in rating points, which is a very direct way to weigh infrastructure changes.

Tom: That’s the big implication for planners, because they can actually price an infrastructure change based on how much perceived safety it’s adding. Physical lane separation is worth something concrete, like point zero nine seven rating points.

Jane: It moves the conversation away from just counting how many people die in accidents and gives them a way to measure the subjective experience of safety that people actually feel while riding.

Meng: I still have some questions about how fast that scoring happens on a CPU when you're running this continuously for thousands of riders. Is it truly fast enough for real-time feedback?

Tom: The scoring part itself is really quick, they say it takes only point zero four milliseconds on a cpu, which makes it pretty light on compute once the attributes are already extracted.

Lu: But we have to remember that the bottleneck might be getting those initial structured attributes from the video stream at the edge in real time. That’s where you really test if this is robust beyond just looking at a few annotated frames.

Jane: So, what we’re seeing here is a system that turns raw, complex video into simple numbers that planners can actually use to make decisions about street design and safety improvements.

Tom: It's moving the tech from just analyzing data after the fact to actively helping people improve their environment while they are in it.

Lu: And while they show how effective this specific setup is, we also have to keep an eye on how these attribute extractions hold up when video quality gets bad because of rain or poor lighting.

Jane: That’s a fair point. So next time we look at these kinds of systems, we'll be asking what happens when the input data isn't perfect.

The paper's improvements: Tom: So, looking at how they suggest improving this system, it’s really about making the edge deployment smoother and more robust for real-world use.

Jane: They are focusing on how to make that attribute extraction part even better so it doesn't break when the video quality gets messy on a bike.

Lu: One improvement is taking that single vision language model prompt and making sure it fits within nine point two gigabytes of GPU memory, which makes running it at the network edge much more feasible.

Meng: That’s exactly what I need to see if we want to ship this thing to actual consumer devices, so fitting the model size is a huge practical step.

Tom: And they are also improving the scoring side by making sure you get those per-feature contributions in rating points, which gives planners even more direct input for their cost-benefit analysis.

Jane: It’s about taking the EBM’s output and making it super clear so that when a planner looks at a safety score, they immediately know exactly which physical feature is driving that number up or down.

Lu: They showed how to make the system more robust by proving it still works even under tighter protocols when they use spatial blocking, which means the model isn't just lucky on clean data.

Meng: What I’m curious about is how they handle those specific feature sets—the cycling lane, construction site, traffic density—if you want to swap them out for something else entirely. Does it still work?

Tom: They’ve done a comparison showing that the performance gap between different vision language models isn't huge for the attributes they are measuring, which is good because it means you have more options in choosing what to use.

Jane: It seems like the authors are really pushing for a system where planners can weigh candidate interventions directly based on perceived risk points rather than relying only on historical accident statistics.

Lu: They're also looking at the continuous monitoring aspect, how to serve thousands of riders without draining their batteries or using too much bandwidth over long distances.

Meng: That continuous monitoring is the hard part for me; if we’re talking about serving a whole fleet, the power consumption and data uplink cost become major deployment hurdles.

Tom: The paper suggests that by re-encoding the video to a twenty second sampling interval, you can serve one thousand riders at fifty four megabit per second, which cuts the payload down twenty three times.

Jane: So they’re looking at practical ways to make this thing continuous so it actually helps with real-time route planning and safety awareness.

Lu: It’s a solid foundation, but we still need to figure out how these improvements hold up when you introduce more complex environments, like extreme weather conditions.

Meng: I’m worried about that; if it can’t handle rain or poor lighting on the camera input, then the whole continuous monitoring idea falls apart in the real world.

Conclusion: Tom: So we’re wrapping up our look at "Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception," and the main point is that we can now use crowdsourced video to give city planners a concrete, explainable measure of cyclist safety perception.

Jane: Exactly. It moves the focus from just counting crashes to understanding what people actually feel about their environment while riding, using data that planners can directly use in their cost-benefit analysis.

Lu: The idea is that by putting the heavy processing at the network edge and using an explainable boosting machine to score it, we get a way to weigh infrastructure changes directly based on how much perceived safety they’re adding.

Meng: I still have some lingering thoughts about making sure this doesn't just sit there in theory; if we want to ship this, we need proof it holds up when the input video quality gets bad because of rain or poor lighting.

Lalam: The paper is quite sure that this specific setup, using a single prompt for attribute extraction and an EBM for scoring, gives us very reliable results on those London routes they tested.

Tom: It’s a solid proof of concept for how crowdsourcing can provide actionable data on cyclist safety perception along routes without needing massive centralized cloud processing.

Jane: It really shows how this technology is becoming a tool for people who want to see the subjective experience of safety scaled into something they can use in decision-making.

Lu: We’ve seen how the system handles attribute fusion with road infrastructure and kinematics, and it seems like a very strong way to integrate different data types into one score.

Meng: For deployment, the main thing we still need to nail down is that scoring speed on consumer hardware when you're pushing it out to thousands of devices simultaneously.

Tom: That’s the next big engineering challenge we have with this pipeline; balancing the accuracy of those video attributes with the strict latency requirements for real-time feedback.

Jane: So, while this paper shows a very clear path forward for using vision language models on edge devices, we still have to keep digging into those practical deployment hurdles and data robustness issues.

Lu: It’s an interesting step because it simplifies attribute extraction by using one prompt instead of needing multiple specialized detection pipelines, which is a big win for model complexity.

Meng: I'm looking forward to seeing how the authors tackle those future work points regarding handling degraded input quality, because that’s where the real world gets messy.

Lalam: I’m confident that this approach provides a very reliable way to get these subjective perceptions into a measurable format for urban planning and safety initiatives.

Tom: That’s what we have here from "Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception," showing how we can use edge computing to make crowdsourced video actually useful.

Jane: It’s a great example of how AI is moving beyond just prediction and into helping people understand their environment in a more nuanced way.

Hongsen Cao, Chia-Yen Chiang, Mona Jaber, Meihui Wang, Yasmin Fathy

School of EECS Queen Mary University of London School of Public Health Imperial College London Department of Engineering University of Cambridge

cs.CV, eess.IV

Submitted: 2026-11-05

Updated: 2026-11-05

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 94/100

The gist: Cyclist fatality rates remain high despite policy support for active travel, and autonomous vehicles add complexity to mixed traffic The EBM regressor achieves 72.2% binary accuracy and MAE 1.011 on…

Key concepts

Edge-Assisted Crowdsensing
This pipeline processes video data directly at the network edge, meaning processing happens locally rather than sending all raw footage to the cloud. This allows for real-time monitoring and reduces bandwidth needs by handling data close to where it is captured.
Single Vision–Language Model (VLM) Prompt
Instead of using many specialized AI models for different tasks, a single text prompt is used to turn each video frame into structured traffic attributes. This simplifies the system by allowing one model to extract multiple pieces of information simultaneously from the visual input.
Explainable Boosting Machine (EBM)
The EBM is an AI scoring tool that predicts a safety score (1-9) based on extracted features like lane separation or construction sites. Crucially, it shows exactly how much each feature contributes to the final score, allowing users to understand why a specific safety rating was assigned.

Terminology

Summary

Cyclist fatality rates remain high despite policy support for active travel, and autonomous vehicles add complexity to mixed traffic The EBM regressor achieves 72. p. 0

Where the paper says it

Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception Hongsen Cao School of EECS Queen Mary University of London London, UK hong. uk Chia-Yen Chiang School of EECS Queen Mary University of London London, UK Mona Jaber School of EECS Queen Mary University of London London, UK Meihui Wang School of Public Health Imperial College London London, UK Yasmin Fathy Department of Engineering University of Cambridge Cambridge, UK Abstract—Cyclist fatality rates remain high despite policy support for active travel, and autonomous vehicles add complexity to mixed traffic.

2% binary accuracy and MAE 1. p. 5
Where the paper says it

The EBM regressor achieves 72. 2% binary accuracy and MAE 1.

011 on a 1–9 scale This system presents an edge-assisted crowdsensing pipeline that turns handlebar video into a structured sketch of each scene at the network edge, fuses it with OSM road infrastructure and rider kinematics, and scores perceived safety with an explainable boosting machine

System Overview

The paper introduces an edge-assisted crowdsensing system designed to monitor cyclist safety perception along a route by processing handlebar camera video at the network edge. This pipeline involves three main stages: capture on the bicycle, attribute extraction at the network edge, and scoring wherever the attribute vector is consumed. p. 1

Where the paper says it

METHODOLOGY Fig. 1 summarises the pipeline in three stages: capture on the bicycle, attribute extraction at the network edge, and scoring wherever the attribute vector is consumed.

The core innovation is using a single vision–language model (VLM) prompt to turn each frame into structured traffic attributes. p. 0
Where the paper says it

Neither crash statistics nor multi-model detection pipelines scale to continuous monitoring of perceived safety along a route. We present an edge-assisted crowdsensing system: handlebar cameras supply street-level frames, a vision–language model (VLM) turns each frame into structured traffic attributes through a single prompt, and an explainable boosting machine (EBM) combines them with OpenStreetMap road infrastructure and on-bike kinematics to predict a crowdsourced 1–9 safety score.

Attribute Extraction and Fusion

The attribute extraction stage utilizes a VLM to convert frames into structured traffic attributes through a single prompt, replacing specialized multi-model detection pipelines. The VLM extracts 14 attributes, with five entering the models: cycling lane, lane separation, construction site, immediate crash risk from traffic and high traffic density. p. 2

Where the paper says it

JSON response, and the model returned 14 attributes. Five enter the models: cycling lane, lane separation, construction site, immediate crash risk from traffic and high traffic density.

These attributes are then fused with OpenStreetMap (OSM) road infrastructure and on-bike kinematics. p. 0
Where the paper says it

Neither crash statistics nor multi-model detection pipelines scale to continuous monitoring of perceived safety along a route. We present an edge-assisted crowdsensing system: handlebar cameras supply street-level frames, a vision–language model (VLM) turns each frame into structured traffic attributes through a single prompt, and an explainable boosting machine (EBM) combines them with OpenStreetMap road infrastructure and on-bike kinematics to predict a crowdsourced 1–9 safety score.

Kinematic features include three-dimensional speed (speed3d) and acceleration (accl3d) derived from GoPro inertial metadata. p. 2
Where the paper says it

Kinematic features. GoPro inertial metadata provides three-dimensional speed (speed3d) and acceleration (accl3d), capturing dynamics the frames cannot show.

Scoring Mechanism

An explainable boosting machine (EBM) is employed as the scoring stage, running in 0.4 ms on a CPU. p. 0

Where the paper says it

49 s per frame and scoring takes 0. 4 ms on a CPU.

The EBM regressor achieves a mean absolute error (MAE) of 1. p. 3
Where the paper says it

Regression. The EBM regressor achieves a mean absolute error (MAE) of 1.

011 and a root-mean-square error of 1. p. 3
Where the paper says it

The EBM regressor achieves a mean absolute error (MAE) of 1. 011 and a root-mean-square error of 1.

303 on the 1–9 scale. p. 3
Where the paper says it

011 and a root-mean-square error of 1. 303 on the 1–9 scale, so the average prediction error is about one point.

Each term in the EBM equation plots as a shape function that shows how a feature raises or lowers the predicted score, allowing for direct attribution. p. 2
Where the paper says it

Each term in Eq. (2) plots as a shape function that shows how a feature raises or lowers the predicted score, and the attribution is read directly from the fitted terms.

Performance and Practicality

On 1,362 annotated frames from two London routes, the EBM regressor reaches 72.2% binary accuracy and a mean absolute error of 1. p. 0

Where the paper says it

On 1,362 annotated frames from two London routes, the EBM regressor reaches 72. 2% binary accuracy and a mean absolute error of 1.

011 under five-fold cross-validation. p. 0
Where the paper says it

2% binary accuracy and a mean absolute error of 1. 011 under five-fold cross-validation.

The system demonstrates robustness, as no model pair separates under a tighter protocol when using spatial blocking. p. 3
Where the paper says it

5%) and random forest (73. Section IV-C shows that no model pair separates under a tighter protocol.

Furthermore, re-encoding the frame to a 20 s sampling interval serves 1,000 riders at 54 Mbit/s at a cost of 1. p. 0
Where the paper says it

4695), an attribute audit, and a sensitivity analysis linking attribute error to accuracy. With re-encoding, a 20 s sampling interval serves 1,000 riders at 54 Mbit/s at a cost of 1.

4 accuracy points, making continuous monitoring practical. p. 0
Where the paper says it

With re-encoding, a 20 s sampling interval serves 1,000 riders at 54 Mbit/s at a cost of 1. 4 accuracy points, making continuous monitoring practical.

Deployment Considerations

The deployment budget shows that scoring a frame with the trained EBM takes 0.4 ms on CPU, while the fastest VLM configuration needs 2. p. 4

Where the paper says it

Scoring a frame with the trained EBM takes 0. 4 ms on CPU; the fastest VLM configuration needs 2.

49 s on an A100-80GB GPU. p. 4
Where the paper says it

4 ms on CPU; the fastest VLM configuration needs 2. 49 s on an A100-80GB, a factor of roughly 6,000.

The payload is reduced by 23× through re-encoding to a 1,024 px long edge at JPEG quality 85, bringing the fleet requirement down to 54 Mbit/s for a 20 s interval. p. 4
Where the paper says it

0 Mbit/s per rider, so a 1,000-rider fleet would demand 5. Re-encoding to a 1024 px long edge at JPEG quality 85 cuts the payload by 23×, and a 20 s interval removes another factor of four, bringing the fleet requirement to 54 Mbit/s.

The EBM regressor’s unsafe-class recall (0. p. 3
Where the paper says it

Section IV-C shows that no model pair separates under a tighter protocol. The EBM regressor’s unsafe-class recall (0.

815) exceeds its safe-class recall (0. p. 3
Where the paper says it

The EBM regressor’s unsafe-class recall (0. 815) exceeds its safe-class recall (0.

623), indicating that the model errs on the side of caution. p. 3
Where the paper says it

815) exceeds its safe-class recall (0. 623), so it errs on the side of caution.

Feature Importance and Interpretation

Individual shape functions dominate pairwise interaction terms in mean absolute contribution, with physically separated cycling lane worth +0.97 points and a construction site costing 0. p. 0

Where the paper says it

97 rating points, a painted lane +0. 70 and a construction site −0.

65 points. The feature combination analysis shows that construction sites show the strongest negative contrast, about −1. p. 2
Where the paper says it

2 shows mean scores for cycling infrastructure combined with three adverse conditions. Construction sites show the strongest negative contrast, about −1.

0 points even where physical separation is present. p. 2
Where the paper says it

Construction sites show the strongest negative contrast, about −1. 0 points even where physical separation is present.

Planners can therefore price an infrastructure change in rating points, and the results enter cost-benefit analysis directly. p. 0
Where the paper says it

70 and a construction site −0. Planners can therefore price an infrastructure change in rating points, and the results enter cost-benefit analysis directly.

The system gives planners a scale of perceived safety on which to weigh candidate interventions. p. 5
Where the paper says it

With scoring some 6,000× cheaper than extraction and a 20 s interval serving a 1,000-rider fleet at 54 Mbit/s, the pipeline is a practical basis for continuous safety monitoring. The system gives planners a scale of perceived safety on which to weigh candidate interventions.

The same edge-assisted pattern extends to other street-level qualities judged by the people who use the street. p. 5
Where the paper says it

The system gives planners a scale of perceived safety on which to weigh candidate interventions. Neither the structured prompt nor the additive model is specific to cycling, so the same edge-assisted pattern extends to other street-level qualities judged by the people who use the street.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Edge-deployable crowdsensing pipeline for structured attribute extraction: The system can turn handlebar video into structured traffic attributes through a single vision language model prompt, replacing specialized multi-model detection pipelines while fitting within 9.2 GB of GPU memory on open-weight models to run at the network edge.

  2. Explainable safety scoring with per-feature attribution: An explainable boosting machine (EBM) can combine extracted attributes with OpenStreetMap road infrastructure and rider kinematics to predict a crowdsourced 1–9 safety score, providing planners with per-feature contributions in rating points.

  3. Infrastructure pricing for cost-benefit analysis: Planners can price an infrastructure change in rating points because the system provides exact attribution, such as physical lane separation is worth +0.97 rating points, allowing results to enter cost-benefit analysis directly.

  4. Scalable continuous monitoring via mobile edge generation: The system allows for continuous route-level monitoring by using re-encoding to cut the payload 23 times and a 20 second sampling interval serving a 1,000 rider fleet at 54 Mbit/s, which costs only 1.4 accuracy points.

  5. Optimized edge compute workload: The system offloads heavy VLM processing to the edge, meaning only a compact structured output (a JSON attribute vector) is transmitted downstream, allowing scoring to take only 0.4 ms on a CPU wherever the attribute vector arrives.

Abstract

Cyclist fatality rates remain high despite policy support for active travel, and autonomous vehicles add complexity to mixed traffic. Neither crash statistics nor multi-model detection pipelines scale to continuous monitoring of perceived safety along a route. We present an edge-assisted crowdsensing system: handlebar cameras supply street-level frames, a vision-language model (VLM) turns each frame into structured traffic attributes through a single prompt, and an explainable boosting machine (EBM) combines them with OpenStreetMap road infrastructure and on-bike kinematics to predict a crowdsourced 1-9 safety score. Extraction belongs at the network edge: the fastest self-hosted model needs 2.49 s per frame and scoring takes 0.4 ms on a CPU. On 1,362 annotated frames from two London routes, the EBM regressor reaches 72.2 percent binary accuracy and a mean absolute error of 1.011 under five-fold cross-validation. It is within one accuracy point of black-box ensembles. Its per-feature shape functions give exact attribution: physical lane separation is worth +0.97 rating points, a painted lane +0.70 and a construction site-0.65. Planners can therefore price an infrastructure change in rating points, and the results enter cost-benefit analysis directly. We also report inter-annotator agreement (alpha = 0.4695), an attribute audit, and a sensitivity analysis linking attribute error to accuracy. With re-encoding, a 20 s sampling interval serves 1,000 riders at 54 Mbit/s at a cost of 1.4 accuracy points, making continuous monitoring practical.

Sources

Related papers