| --- |
| datasets: |
| - pdDuduzin/ESP32-CAM_Autonomous_Line-Following_Car_Dataset |
| tags: |
| - tensorflow-lite |
| - esp32 |
| - computer-vision |
| - line-following |
| - embedded-ml |
| license: apache-2.0 |
| language: |
| - en |
| - pt |
| --- |
| |
| # ESP32-CAM Autonomous Line-Following Car Model |
|
|
| Quantized steering model used by an autonomous line-following car built with an ESP32-CAM and an ESP32 DevKit. |
|
|
| ## Model contract |
|
|
| - Input: `96 x 96 x 1` grayscale image. |
| - Output: one steering value from `-1` (maximum left) to `+1` (maximum right). |
| - Task: supervised regression by behavioral cloning. |
| - Deployment format: fully quantized INT8 TensorFlow Lite Micro model. |
| - Dataset: [ESP32-CAM Autonomous Line-Following Car Dataset](https://huggingface.co/datasets/pdDuduzin/ESP32-CAM_Autonomous_Line-Following_Car_Dataset). |
|
|
| ## Files |
|
|
| - `modelo_linha.tflite`: quantized deployment model. |
| - `modelo_linha.h`: C header embedded in the ESP32-CAM firmware. |
| - `modelo_linha.keras`: original Keras model. |
| - `training_manifest.json`: training configuration, session split and hashes. |
| - `conversion_manifest.json`: quantization configuration, validation metrics and hashes. |
| - `avaliacao_val.json`: validation report. |
| - `historico_treino.png` and `precisao_val.png`: training and validation plots. |
|
|
| ## Validation results |
|
|
| For the INT8 model on the validation split: |
|
|
| - MAE: `0.2585` |
| - MSE: `0.1145` |
| - R²: `0.5372` |
| - Curve direction accuracy: `95.81%` |
|
|
| These offline metrics do not guarantee stable driving. Performance is sensitive to lighting, reflections, track appearance, camera position and conditions that differ from the training data. |
|
|
| ## Deployment |
|
|
| Copy `modelo_linha.h` to `firmware/esp32cam_autonomous/modelo_linha.h`, compile the autonomous firmware and upload it to the ESP32-CAM. The header in this release has SHA-256 `f835b35dae58bbae79324b01cc249774e020db102483add12ce391a6d3fae751`. |
