SPARDHA: A Mobile-First Framework for on-Device Pothole Detection, Route-Aware Alerts, and Automated Reporting of Road Hazards To Civic Authorities

Uncategorized

Authors: Dr. Rudramurthy V C, T Dhanush, Yagnesh L, Vikas Reddy

Abstract: Potholes and unmarked speed breakers are making rides uncomfortable, wear vehicles out faster and create a genuine hazard for road safety, broadly speaking; indeed, manual road inspections are slow, and complaints handed in by citizens often arrive without location data or any supporting indication, in actual application. Till then, most people already carry smartphones whose motion sensors are capable of sensing road-surface challenges, in real-world scenarios. A React Native app is collecting data from the phone's accelerometer and gyroscope, splits the stream into 20 x 8 windows and runs a TensorFlow Lite model on the device to label each window as NORMAL, POTHOLE or SPEED_BREAKER, broadly speaking. Turning phone vibrations into the usable road information, though, isn't just a matter of building a classifier, like the framework must record and analyse data, attach a location to every detection, warn the rider in time, combine repeated reports of the same issue, remain usable by the public, and send summaries to the right authorities without anyone filing a complaint by hand; this paper is offering SPARDHA, a system that's tackling all of these stages are ending to end. Every trip is, indeed, typically written to a CSV file for later study or retraining, and when a pothole is detected, the rider receives an alert before getting the hazard; the detection is then stored locally in SQLite and uploaded to a server whenever the network becomes available. The server, built on FastAPI, Kafka and PostGIS, accepts a detection only when the model score is at least 0.70, merges same-type reports within 15 metres, and sends a summary of supported potholes to the responsible authorities so that repairs can be planned without waiting for complaints. The whole system is running on Google Cloud, which is making it accessible to anyone with an Android phone; in fact, testing on an Android release build confirmed that the model loads, generates real-time predictions, raises rider alerts, connects to the API, an application programming interface, an application programming interface, and filters trips by route. The paper closes by setting out the field evaluation that any claim about precision, speed or reliability will (to some degree) have to pass.

DOI: https://doi.org/10.5281/zenodo.24341321

× How can I help you?