How elfilaoussama/advanced_driver_vision_assistance_with_near_collision_estimation_system works
Main components
• Client/Vehicle Side: Represents the thin client (on-board the vehicle) capturing image data and sending it to the server.
• API / Server (app.py): Acts as the central entry point, receiving data, routing requests, and coordinating the processing pipelines.
– DETR (object detection)
– GLPN (depth estimation)
– LSTM (historical data prediction)
• Data Storage & Configuration:
– The “data” directory holds pre-trained models and scaling objects.
– “config.py” is present for system-wide configuration settings.
• Utilities and Extensions:
– “utils” for JSON output formatting, Cython-based processing optimizations, and visualizations.
• Training & Testing Components:
– “trainings_tests” contains Jupyter notebooks and test data for model training, evaluation, and refining dataset pipelines.
• Deployment:
– The Dockerfile and configuration files (.idea, .vscode) point to containerization and development environment settings.
Relationships and data flow
• The server, upon receiving data, dispatches the image data to the different model modules:
– First, detected objects can be obtained from the DETR module.
– Then, depth estimation via the GLPN module.
– And finally, historical trajectory or prediction through the LSTM module.
• The configuration (config.py) should be shown as providing runtime settings that influence one or more of the modules.
• Utilities (under utils) should be depicted as auxiliary services employed by the server for data formatting, visualization, or optimized processing (via Cython).
Important architectural patterns/design principles
• Client-Server Architecture: Highlight the separation between the thin client (vehicle) and the cloud-based processing system.
• Modular Model Processing: Clearly separate individual processing blocks (DETR, GLPN, LSTM) to emphasize independent responsibilities and easier testing/updating.
• Data Pipeline & Integration: Indicate data ingestion, pre-processing (e.g., scaling using the models in the “data” folder), and post-processing (JSON output via utils/JSON_output.py or visualization).
• Containerization: Mark where Docker is involved, suggesting that the entire server (and perhaps training sessions) might run within containerized environments for scalability and consistency.
• Legends or annotations that explain each component’s role and the technologies involved (mention of PyTorch for LSTM/deep learning models, XGBoost for potential additional modeling, Cython for performance optimizations, etc.).