Melissa Adaptive
Signaling Technology
I'm Pranil Thatavarthy. I live in Melissa, Texas, where a line of cars waiting to turn can hide the traffic coming toward you. I'm building MAST to help drivers see that traffic sooner.
Prototype in development · October 2026
Why this intersection
At SH 121 and Milrany Lane, cars in the opposite left-turn lane can hide the traffic coming straight through. MAST asks whether an affordable radar sensor could give drivers more information before they turn.
See the research →The bigger project
The original MAST proposal used YOLOv11 to recognize vehicles from a camera feed. The idea was to adjust signal timing as traffic changed, including giving priority to emergency vehicles.
That work began with COCO and Open Images data. Local traffic patterns are a separate part of the timing research.
The radar prototype focuses on a different problem: the car you cannot see. It measures approaching traffic and estimates how soon each vehicle will reach the intersection.
The radar watches oncoming traffic from the opposite direction. Its advisory appears in the traffic light facing the driver waiting to turn left, so there is no separate roadside display to look for.
Look inside the prototype →Explore the prototype
Scroll over the model or choose a stage below.
Drag to look around.
Loading the 3D model…
How MAST has developed
2025 → 2026
MAST started as a team entry to the Presidential AI Challenge, exploring camera-based adaptive signal timing. We did not advance past regionals. In summer 2026, I returned to the project and focused on the blocked view during a left turn.
NCTCOG · Gregory Masota
Gregory Masota at NCTCOG helped me obtain crash summaries and INRIX turning-movement estimates. The crash records showed 13 left-turn crashes at SH 121 and Milrany Lane from 2021 through 2025. Six involved an injury.
I then used TxDOT's CRIS tool to look at intersections in Prosper, Plano, and Frisco. That became the comparison I brought to the Melissa meeting.
Read the findings →Melissa Public Works
I met with Matthew Dill and Frankie Galvez to discuss the crash pattern and explain what I was trying to build. We discussed who manages the intersection and which engineering questions I needed to answer.
They confirmed that TxDOT manages the signals on SH 121 and connected me with Jennifer Vorster. They also suggested speaking with traffic engineers and companies that build traffic systems.
TxDOT · Jennifer Vorster
Jennifer Vorster, a Collin County Area Engineer at TxDOT, helped me think through what drivers would actually notice. She raised questions about pedestrians, colorblind drivers, and whether the display could distract someone who is trying to turn.
She pointed me toward the Texas A&M Transportation Institute and asked me to reconnect after that conversation. I am continuing to work through the design with engineering guidance.
September 28, 2026
I met with Srinivasa Sunkari and David Florence at the Texas A&M Transportation Institute to discuss the prototype and how to test it in local traffic-cabinet circuit settings.
We agreed to plan controlled local tests to see how the sensor behaves around real traffic-cabinet circuitry. Those results will help me decide what needs to change.
Crash records
Selected left-turn crashes involved injury
at SH 121 and Milrany Lane, 2021–2025.
NCTCOG supplied two crash summaries for SH 121 and Milrany Lane. Together, they contained 43 crashes from 2021 through 2025, including 13 left-turn crashes. Of those 13, three involved suspected minor injury, three involved possible injury, and seven involved property damage only.
| Location | Crashes | With injury |
|---|---|---|
| MelissaSH 121 & Milrany Ln | 13 | 46% |
| ProsperCoit Rd & US 380 | 15 | 47% |
| PlanoParker Rd & Custer Rd | 54 | 50% |
| FriscoEldorado Pkwy & FM 423 | 27 | 41% |
For the comparison sites, I filtered TxDOT CRIS records for “failed to yield right of way, turning left,” then reviewed the Plano and Frisco results for the oncoming-traffic pattern. I left Anna out because only three records matched. The Frisco figure includes one fatal crash. The records describe the crashes and their severity; they do not establish which drivers had a blocked view.
TxDOT CRIS public query ↗Sensor range
The faster a car approaches, the farther away it needs to be detected for the same time window. The OPS243-C has a reported range ceiling of 60 meters. At 55 mph, a five-second window needs about 123 meters, so the current plan places the radar farther up the road. The exact mounting location still needs a site survey.
0 mSensor ceiling: 60 m175 m
About 63 meters beyond the sensor's reported range.
Inside the build
I chose the OmniPreSense OPS243-C because it reports distance as well as speed and direction. It sends those measurements to the Raspberry Pi over USB. Its 20-degree beam needs to be aimed at the approaching lanes.
The software divides each approaching vehicle's distance by its speed, then watches the shortest arrival time. The radar part uses these measurements directly; it does not need a trained image-recognition model.
Moving the radar upstream gives earlier readings, but the car eventually passes it. The software keeps the last confirmed track and projects the arrival time instead of assuming the road is clear when the reading disappears.
The first version uses one radar and retained tracks. A second sensor closer to the intersection is an option I want to evaluate for additional coverage. Braking, lane changes, and multiple approaching vehicles all need testing.
The design uses four horizontal left arrows: red, yellow, a flashing advisory arrow, and green. The third arrow can flash yellow, red, or green according to the estimated gap in opposing traffic. The sensor watches the opposite approach and the display faces the driver waiting to turn left. Drivers still follow the signal and yield to traffic.
Try the display in 3D →Where we are now
The Raspberry Pi is set up, the OPS243-C is connected over USB, and I have received raw Doppler readings. The build log records 31 passing software tests for tracking, sensor health, and edge cases. I am now checking the software against the actual sensor output.
Match the software to the radar's actual output, including its direction convention: positive values mean a vehicle is approaching. Check how multiple vehicles appear in the data.
A vehicle passing the radar can still be approaching the turn. Test how long the software retains that track, and how it responds when a reading goes missing or the sensor stops reporting.
Follow the plan discussed with Srinivasa Sunkari and David Florence: evaluate the sensor in a local traffic-cabinet circuit setting and take the measurements back for review.
Survey a possible upstream mounting location. Refine the flashing third arrow in the horizontal signal, including how it reads for colorblind drivers. Pedestrian detection is another question I still need to address.
The reason for MAST
I'm a high school student in Melissa, Texas. I started MAST because this is a problem close to home. My goal is to find out whether a low-cost radar advisory can help, and to be clear about what the tests show as I go.
Explore the prototype ↑