An open research program on seeing traffic, simplifying road networks, and controlling signals, with every result traceable to its data.
Soheil Fallahzadeh, M.Sc. student in computer vision and algorithms, Yazd University.
This work extends my M.Sc. thesis on machine vision, deep learning and fuzzy logic for intersection control toward network-level emergency navigation.
Every routing or signal decision downstream is only as good as what the camera reports. Most of my work sits at that boundary: measuring what a detector gets wrong, and passing that uncertainty on to the controller instead of hiding it.
A reliability score computed without ground truth, from image quality and detector health.
0.938 AUC on held-out synthetic streams
Checked on 12 human-labelled UA-DETRAC clips (sunny, cloudy, rainy, night) with an unmodified YOLOv4-tiny: combined AUC 0.973 on a temporal split, median ECE 0.036 when a clip is held out (worst 0.116). Calibration did not transfer between two pseudo-labelled clips (ECE 0.12 to 0.23), and that is reported.
Queue length behind a truck, with an uncertainty range, from an oblique camera.
−2.84 → −0.63 m bias under occlusion
Flashing-light cue, HMM map matching and heading, delivered as one event a signal controller can act on.
0.974 road matching (nearest-road: 0.900), 1.1 ms per frame
Graph-gated vehicle re-identification and network tomography on sparse cameras.
0.875 association precision (appearance only: 0.558)
Replicated on 10 seeds and a second grid with bootstrap intervals: the gain from graph gating holds, the gain on unobserved edges does not.
The graph work and the vision work are one program: the network has to be simple enough to reason about, and the sensors feeding it have to be honest about their own limits.
An emergency vehicle needs three things from a city: eyes that know when they can’t be trusted, a road network simple enough to reason about quickly, and signals that react before a junction locks up. Each layer is its own repository, tested on its own terms.
Cameras see traffic, but they get fooled by rain, glare and parked trucks. These five repositories measure how much to trust what they report.
Perception reliability monitor. Scores a camera’s health without ground truth and tells a controller when to fall back to a fixed plan.
0.938 AUC predicting when detection is reliable (image quality alone: 0.873)
Raises the alarm within 3 to 8 seconds when the detector actually fails, and stays quiet when it copes.
| Features | AUC | Calibration error |
|---|---|---|
| Image quality only | 0.873 | 0.070 |
| Detector health only | 0.901 | 0.078 |
| Combined | 0.938 | 0.049 |
14,901 test frames, 6 disjoint seeds, synthetic scenes.
Occlusion-aware queue estimation. Reasons about what a roadside camera cannot see and reports queue length with an uncertainty range.
Bias cut from −2.84 m to −0.63 m under occlusion
Picks the busiest lane correctly more often: 77.7% of the time, up from 71.5%.
| Variant | MAE (m) | Bias (m) | Lane top-1 |
|---|---|---|---|
| Count the boxes | 5.22 | −1.16 | 0.778 |
| Visible chain only | 4.48 | −2.84 | 0.715 |
| Occlusion-aware | 4.63 | −0.63 | 0.777 |
18 scenarios, bootstrap intervals. Overall error is not significantly better; the gain is where occlusion is heavy. Interval coverage is about 0.74 against a nominal 0.95.
Emergency-vehicle perception for preemption. Turns camera tracks into one event a signal controller can act on: road edge, distance, arrival time, confidence.
Road matching accuracy 97.4% (nearest-road baseline: 90.0%)
Median 1.1 ms per frame on one CPU core. In simulation, preemption cut ambulance passage from 34.4 s to about 20 s.
| Cues | Detection rate (95% CI) | False alarms per track |
|---|---|---|
| Detector only | 0.68 [0.58, 0.76] | 0.0133 |
| Flashing light only | 0.65 [0.55, 0.74] | 0.0017 |
| Fused | 0.77 [0.68, 0.84] | 0.0017 |
100 emergency and 600 ordinary tracks. Intervals overlap, so the fusion gain is not established at this sample size. Perception is emulated, not real footage.
A real road network has tens of thousands of segments. A spanner keeps a small subset that still guarantees no trip gets much longer.
Geometric spanners lab. A clean, reproducible baseline for the classical greedy t-spanner.
990 streets reduced to 77, 92.22% pruned, detours capped at 1.4×
| Property | Complete graph | Greedy spanner |
|---|---|---|
| Edges (45 points) | 990 | 77 |
| Stretch | 1.00 | at most 1.40 |
| Density | O(n²) | O(n) |
Geo-SmartSpanner. A graph neural network predicts which roads to drop, and a repair step restores the guarantee on directed city graphs.
Up to 42.5× faster than greedy on Rome (42,788 nodes)
Held-out cities showed no zero-shot loss in quality.
| City | Nodes | Greedy (s) | Learned (s) | Speedup |
|---|---|---|---|---|
| Manhattan | 4,540 | 0.742 | 0.110 | 6.74× |
| Eindhoven | 7,881 | 1.948 | 0.206 | 9.45× |
| Paris | 8,236 | 3.256 | 0.193 | 16.89× |
| Rome | 42,788 | 46.626 | 1.097 | 42.51× |
Mean of 5 seeds, stretch bound 1.5. A small violation rate (0.011%) remains in Eindhoven. The guarantee is per removed edge, not per pair.
Delay-bounded spanner networks. One road backbone certified to keep travel-time detours bounded at every hour of the day.
A proof that local checks give an all-pairs, all-24-hours guarantee
Tested on 12 cities across six countries, with an independent verifier that is itself tested against planted errors.
| Evidence | Where |
|---|---|
| Theorem 1 and algorithm | README, sections 1 and 2 |
| Eight independent result checks | verify_results.py |
| Raw per-pair dilation | results/per_pair_dilation.csv |
Congestion is a parametric model, not fitted to telemetry. The 12-city result supports a directional, not definitive, claim.
At the junction, a controller has seconds to decide. It must move traffic without blocking the intersection itself.
Anti-gridlock signal control. Max-pressure control with a fuzzy throttle that backs off when the road ahead is filling.
Ambulance delay 143.3 s to 125.0 s with preemption (p = 0.049)
Under chronic overload, 342.6 s average delay against 380.3 s for plain max-pressure, and about 10% less CO₂.
| Policy, chronic overload | Delay (s) | Throughput |
|---|---|---|
| Fixed time (Webster) | 306.2 | 1374 |
| Vanilla max-pressure | 380.3 | 1374 |
| Fuzzy anti-spillback | 342.6 | 1373 |
10 paired seeds, Holm–Bonferroni corrected. When the bottleneck is loose, the controller is worse than both baselines. A silent unit-scaling bug found mid-project is disclosed in full in the repository.
The layers are designed to meet at a single intersection. How they compose is the next study.
A result that fails is still a result. These are reported in the repositories, not buried.
The monitor detects failures reliably, but in simulation the controller stayed within 0.3 s of the oracle anyway.
perception-reliability-monitorThe benefit was flat from 30 m to 150 m at normal demand. It grows once demand doubles.
ev-perception-for-preemptionOn 437,623 nodes it stayed correct but removed only 0.09% of edges.
Research_Geometric_ML_OptimizationAt loose bottlenecks it is slower than both baselines. The cost shifts congestion to the arterial, it doesn’t remove it.
anti-gridlock-signal-controlThe nine repositories each answer a narrow question. These are the questions they leave open, and where I want to go next.
Reliability and queue-uncertainty scores already exist as outputs. No controller consumes them yet.
The certificate assumes known edge costs. Real costs are estimated, and the estimates can be wrong.
Preemption is tested on one approach and one signal. A corridor needs many intersections to agree.
All nine repositories are MIT-licensed. The perception and control repositories use fixed seeds, no external datasets, and list the exact commands to regenerate every table. Most carry a claims-and-evidence table that marks each claim as supported, partial or not supported.