| Problem Statement 3 Real-Time Road Anomaly Detection from Dashcam Footage on Raspberry Pi |
| Objective |
|
|
| Build an edge AI application on Raspberry Pi that processes dashcam footage in real-time to detect and log road anomalies such as potholes and unexpected obstacles. |
| Project Description |
|
|
| Students will choose a lightweight object detector (e.g., MobileNet-SSD, YOLOv5s), convert it to an edge-optimized format (TensorFlow Lite / ONNX Runtime / ExecuTorch), and integrate it with an OpenCV video pipeline. Detected anomalies should trigger timestamped logs or saved clips. |
| Key Requirements |
|
|
| Hardware: |
| Raspberry Pi 5 or 4. |
| Raspberry Pi Camera Module v2 or USB webcam. |
| High-write-speed microSD card. |
| Where possible, aim to use the CPU without additional accelerators/hats. Solutions that are well-optimised through use of Quantisation, KleidiAI, and appropriate model selection - and therefore able to run entirely on CPU - are of great interest. |
| Software: |
| Raspberry Pi OS. |
| Python, OpenCV. |
| TensorFlow Lite / ONNX Runtime / ExecuTorch with a pre-trained, quantized detection model. |
| |
| Performance Targets |
|
|
| ≥5 FPS near-real-time inference. |
| High precision to reduce false positives in logging. |
| Robust under varying lighting conditions. |
| |
| Deliverables |
|
|
| Source code for video processing and inference pipeline. |
| Optimized deployed model file (.tflite / .onnx). |
| Demo video with anomaly detection on sample footage. |
| Report on model choice, optimization and performance. |
| |
| Learning Outcomes |
|
|
| Optimizing and deploying neural networks for edge video analytics. |
| Experience with embedded vision pipelines. |
| Understanding accuracy vs speed vs compute trade-offs on Arm platforms. |
| |
| Mentoring session schedule and details |
| PS# Date Time Meeting Link |
| 3 9th Feb 3:30 – 4 PM Zoom link |
| Meeting ID: 957 4790 4145 |
| Passcode: 521992 |
|
|
