This is a breakdown of a project I built to detect rail fractures using low-cost hardware (ESP32-CAM) instead of expensive industrial sensors.
The Constraints:
Hardware: ESP32 (240MHz, 4MB PSRAM).
Model: MobileNetV2 (Originally 130MB Float32).
Goal: Offline inference on the edge.
The Solution: I used Post-Training Quantization (Int8) to shrink the model to 10.9MB. The system achieves ~1.5 FPS (600ms latency), which is slow for video but sufficient for maintenance trolleys moving at low speeds. It communicates alerts via MQTT when connectivity is available.
Happy to answer questions about the quantization pipeline or the hardware setup!
hemanthmuralik•1mo ago
This is a breakdown of a project I built to detect rail fractures using low-cost hardware (ESP32-CAM) instead of expensive industrial sensors.
The Constraints:
Hardware: ESP32 (240MHz, 4MB PSRAM).
Model: MobileNetV2 (Originally 130MB Float32).
Goal: Offline inference on the edge.
The Solution: I used Post-Training Quantization (Int8) to shrink the model to 10.9MB. The system achieves ~1.5 FPS (600ms latency), which is slow for video but sufficient for maintenance trolleys moving at low speeds. It communicates alerts via MQTT when connectivity is available.
Happy to answer questions about the quantization pipeline or the hardware setup!