Cars can be tracked through surveillance cameras nobody gave access to

What encrypted camera traffic can reveal without showing video

Encrypted video still leaves a clear fingerprint in the network, because each frame is cut into packets whose sizes and arrival times are visible without opening the picture. Those packet clues change when a car enters view, because compression compares each new picture with the previous one. A moving car creates more detail than a quiet street. That difference survives encryption in a detector that watches only packet arrival times and sizes. It never reads the picture or the car’s plate. The detector was trained on 2,607 encrypted camera traces from 237 simulated runs over eleven cameras. It was then judged on a held out set of 242 sequences. That test gave an F1 score of 0.989. So the score combines two rates, the share of real passages caught and the share of reported passages that are true. Its measured errors were a missed event rate of zero and a false event rate of 0.0225 per real passage. Those errors mean each clue only says that some vehicle passed near a known camera at roughly a given time. It does not say which car, what color it was, or where the driver went next. The measured timing error has an average shift of 0.022 seconds and a spread with a standard deviation of 0.109 seconds. That timing error is the kind of weak clue fed into the tracker.

How weak anonymous clues fill the gaps between accessible cameras

Between the cameras a tracker can actually use, a car can disappear for long enough that its position becomes almost guesswork. The accessible cameras are few, and they usually sit at chosen points such as intersections or perimeter gates. When one of them sees a vehicle, it gives a strong clue: a time, a place, and an identity. Once the next camera has not yet seen it, the tracker must carry that uncertainty forward through streets where no direct view exists.

Anonymous passage events can fill part of that gap. Each event gives only an approximate timestamp for the passage near a known third-party camera. That timestamp carries an average offset of 0.022 seconds, and its typical spread is 0.109 seconds. The clue does not say which car was there, and it may be late, early, missed, or wrong. Still, the clue is useful because it lands inside a blind stretch. If a tracker knows the car left one accessible camera heading north and later appeared at another, an anonymous event near a side street can rule out routes that never pass close to that point. The evidence is weak, but it is not empty.

So a controlled simulation of 120 vehicle trips across the CARLA Town05 road network made that value measurable. The simulated region had 12 sparse direct sensors around its edge and 15 interior cameras whose observations were treated as anonymous passage events. With direct sensors alone, the average position error was 92.3 meters, and 35.8 percent of trips ended with errors above 100 meters, meaning the estimated path had left the right neighborhood entirely. Adding noisy anonymous events cut that average error to 36.8 meters, a reduction of 60.1 percent. The severe failures dropped from 35.8 percent to 0.3 percent, close to the near-perfect case where every intermediate event was correct.

The point is not that these clues replace direct cameras. They are too coarse for that. Their value appears in the gaps, where a tracker would otherwise rely on speed and direction alone. A single anonymous passage can be mistaken, but several such events spread across a route give the tracker repeated chances to discard impossible paths. That result is a smoother estimate of where the vehicle was between strong sightings, even when no one has access to the video that produced the clues.

Why road maps matter when a car disappears from view

When a car leaves an accessible camera, the tracker loses its clearest clue. That vehicle may still be moving through streets that no one can watch directly. Its last known position becomes a starting point, and from there many routes seem possible until a road map narrows them. That map is not open space, so cars cannot cross blocks, fly over parks, or drive through buildings. So they must follow connected lanes and plausible travel times. Then the tracker can discard guesses that would require impossible turns or impossible speeds. This matters most during blind gaps, when no new direct sighting arrives. Once blind gaps opened up, a simulation of 120 vehicle trips through a generated town showed the gain. So road constraints cut average position error by 27.6 percent compared with an unconstrained filter using the same clues. That benefit grew as the gap between reliable sightings lengthened. It was about 5.4 meters on shorter gap routes, and it reached about 26.8 meters on longer ones. That pattern points to drift: without a map, guesses wander farther from the real path over time. The result is not that intersections alone explain the gain. Instead, the simulation found no clear link between how many branches a route had and how much the road map helped. Rather, the map mattered because it kept every guess on streets where a car could actually be. During long blind stretches, that simple physical rule did more work than raw motion guessing.

Cars can be tracked through surveillance cameras nobody gave access to

A vehicle can be followed through blind stretches without ever seeing its face, plate, or video feed. What matters is a thin stream of anonymous signals saying only that some car passed near a known camera around a certain time. When those weak events are combined with sparse official sightings and the physical layout of streets, the tracker stops drifting into impossible places.

In controlled tests built from 120 simulated drives through a city-like road network, adding noisy indirect clues cut average position error from 92.3 meters to 36.8 meters. That is a 60.1 percent reduction. Catastrophic losses fell from 35.8 percent of routes to 0.3 percent.

A simple picture makes the idea concrete: a car moves along a road map, official cameras mark only a few points, private cameras add faint dots in between, and the tracker uses street connections to choose the most plausible path. The broader point is that cities already hold far more sensing hardware than any single operator can directly control. Private storefronts, residents, businesses, and other organizations may own cameras that never share raw footage. If they expose only presence events, or if encrypted traffic reveals enough packet timing and size to infer a passage, tracking can extend into gaps where official sensors are absent. The same idea could apply to motion detectors, proximity sensors, or feeds that report movement without identity. That makes the approach useful for urban mobility, public safety, security operations, and defense settings where coverage is expensive and access is fragmented.

The obstacles are real, and they include calibration needs, changing network conditions, and the harder problem of matching anonymous events to moving vehicles. The passage detector was trained on labeled camera traces, and its accuracy may shift when codecs, viewpoints, or wireless conditions change. Tracking many vehicles at once remains difficult because an event does not say which car produced it, so assignment accuracy falls as traffic density rises. In the multi-vehicle tests, matching accuracy fell from 100 percent with one car to 87 percent with two and 63 percent with three. The controlled results show a substantial improvement under the modeled conditions, but they do not prove that real cities can turn scattered private cameras into a working tracking layer. Ownership, privacy, cost, and operational constraints can limit direct access, so the tested pipeline is one possible route rather than a ready citywide system.