Real-Time Object Detection for Autonomous Vehicles

📌 Project Description
This project builds a deep learning–based real-time object detection system for autonomous vehicles. It detects and classifies key road objects—Cars, Pedestrians, Cyclists—to support safe navigation across diverse conditions.

The model is trained on the KITTI Object Detection dataset. In Phase 1, a 100-image cleaned subset established the pipeline. In Phase 2, we scaled to the full KITTI dataset (~12 GB) for comprehensive EDA and feature engineering. In Phase 3, we trained and validated multiple YOLO models with class balancing and augmentation. In Phase 4, we deployed the best model to a Flutter Android app that performs continuous on-device inference.

---

🚀 Project Phases

Phase 1: Data Preparation (✅ Completed)

• Collected and uploaded a 100-image KITTI subset (images + labels).
• Cleaned the dataset:
  • Removed images without labels.
  • Excluded empty label files.
  • Detected and removed duplicate images.
  • Filtered invalid/outlier bounding boxes.
• Converted annotations from KITTI → YOLO format.
• Split into 80% train / 20% val.
• Generated data.yaml for YOLO training.
• Exported a cleaning_report.json of excluded files.
• Packaged the final subset as kitti_yolo_prepared.zip.

---

Phase 2: EDA & Feature Engineering (✅ Completed)

• Scaled analysis to the full KITTI (~12 GB).
• Descriptive statistics on boxes, labels, image distributions.
• Visualized patterns:
  • Class frequencies (cars dominant; pedestrians/cyclists rarer).
  • Bounding-box sizes and aspect ratios.
  • Object-center heatmaps and objects-per-image.
• Addressed imbalance and quality:
  • Encoded labels for YOLO.
  • Normalized boxes relative to image dimensions.
  • Applied scaling/augmentation strategies to improve balance.

---

Phase 3: Model Training & Validation (✅ Completed)

• Configured YOLOv8 training on the prepared dataset.
• Trained multiple variants (e.g., YOLOv8n, YOLOv8s) for comparison.
• Augmentations: mosaic, mixup, flips, HSV jitter, scaling.
• Imbalance handling:
  • Oversampled minority classes (pedestrians, cyclists).
  • (Optionally) weighted loss.
• Tracked metrics: mAP, precision, recall, F1.
• Visualized results:
  • Confusion matrices, PR/RC/F1 curves.
  • Example predictions with bounding boxes on KITTI images.
• Exported trained models and logs:
  • best.pt, best.onnx, best_float32.tflite, best_float16.tflite, best_int8.tflite, best.torchscript, saved_model/, ncnn/.
• Selected the best configuration for deployment (FP16 TFLite recommended on Android).

---

Phase 4: Mobile Deployment (✅ Completed)

• Built a Flutter Android app with:
  • Live camera preview (CameraX) and continuous inference (process every N-th frame for smooth UI).
  • YOLO letterbox pre-processing and reverse mapping to preview coordinates.
  • On-device NMS and overlay rendering (boxes, labels, confidences).
• Android config highlights:
  • compileSdk=36, minSdk=26.
  • Keep .tflite uncompressed for fast mmap loads.
  • Optional TFLite GPU dependency + keep rules when using GPU delegate.
• Produced debug and release APKs.