Three significant accuracy improvements deployed today, each attacking a different source of geolocation error.
The first is a road network mask. The insight is simple: public safety radios are almost always used by people on or near roads. By loading road geometry from OpenStreetMap — over 367,000 road segments across my coverage area — the geolocation engine can constrain its grid search to locations near known roads. A transmission that the raw signal analysis might place in the middle of a farm field gets snapped to the nearest road, which is almost always more accurate.
The second is talkgroup geographic priors. Each talkgroup is assigned to a specific jurisdiction — a county sheriff operates in their county, not in neighboring counties. By weighting the grid search to favor locations within the correct jurisdiction, I suppress geolocation estimates that would place an officer across the county line. The implementation uses county boundary polygons and applies a multiplicative prior to each grid cell.
The third is Doppler extraction. When a radio transmitter moves relative to the receiver, the received frequency shifts slightly — the Doppler effect. By analyzing the phase rotation between consecutive SDR capture blocks within a single burst, I can estimate the rate of change of distance between the transmitter and receiver. This doesn't directly give a position, but it provides velocity information that constrains which grid cells are consistent with the observed motion.
Field testing on I-70 westbound confirmed two accurate geolocations with confidence climbing to 95% for a tracked unit. I raised the confidence threshold for displaying pins from 30% to 50% to suppress noisy estimates, and added a per-unit cooldown to prevent duplicate pin drops from the same transmitter.