Vehicle Safety Detection Pipeline with AWS Lambda

This repository contains the implementation of a vehicle safety detection pipeline using AWS services to detect pedestrians in vehicle camera images, process them with a computer vision model, and store results in a database. This project was developed as part of my preparation for a internship interview, aligning with the role's focus on data science, AI, computer vision, and cloud optimization.

Project Overview

• Purpose: Detect pedestrians in images to identify critical road safety situations, contributing to the evaluation and optimization of vehicle safety systems.
• Goal: Demonstrate skills in cloud engineering, computer vision, database design, and Python programming by building an end-to-end pipeline.
• Architecture: S3 → AWS Lambda Layer → RDS (PostgreSQL).

System Architecture

• Amazon S3: Stores uploaded vehicle camera images, triggering the pipeline.
• AWS Lambda Layer:
  • Processes images using a YOLOv8 ONNX model pre-trained on COCO for pedestrian detection.
  • Custom layer includes dependencies: onnxruntime, Pillow, psycopg2, and numpy.
  • Pre-processing with Pillow (e.g., resizing to 640x640).
  • Post-processing with Non-Max Suppression (NMS) to avoid duplicate pedestrian counts.
• Amazon RDS (PostgreSQL): Stores detection results in the vehicle_safety_data table (image key, timestamp, location, speed, classification, pedestrians detected).
• Hardware: Optimized for x86_64 architecture to match AWS Lambda.

Challenges and Solutions

1. Lambda Layer Size Constraint (250 MB Limit)

• Challenge: The 250 MB limit for Lambda layers made it difficult to include all libraries (onnxruntime, Pillow, psycopg2, numpy) and the 5.47 MB YOLOv8 model.
• Solution:
  • Optimized imports by using minimal versions (e.g., onnxruntime==1.16.3, Pillow==9.5.0) with --no-deps.
  • Chose Pillow over OpenCV for lighter image processing.
  • Used Docker to build a compact layer on Amazon Linux 2, removing .dist-info and .egg-info directories.

2. Building Compatible Libraries for Lambda

• Challenge: Building dependencies for Lambda’s Amazon Linux environment from an ARM-based Python 3.12 setup was incompatible.
• Solution: Utilized Docker containers with Amazon Linux 2 to compile and package dependencies (pip install --platform manylinux2014_x86_64), ensuring compatibility with Lambda’s runtime.

3. Model Selection Under Constraints

• Challenge: Needed a model that fits within the 250 MB limit while providing accurate pedestrian detection.
• Solution: Selected YOLOv8n (lightweight variant) in ONNX format, optimized with float32 inputs to reduce memory usage, fitting the 5.47 MB model into the layer.

4. Enabling Communication Between AWS Services

• Challenge: Setting up secure communication between S3, Lambda, and RDS required proper permissions and VPC configuration.
• Solution:
  • Configured IAM policies with s3:GetObject, s3:ListBucket, and EC2 actions (CreateNetworkInterface, etc.) using specific ARNs (e.g., arn:aws:s3:::vehicle-safety-bucket/*).
  • Set up VPC with subnets and security groups for RDS access, resolving permission errors.

5. Avoiding Duplicate Pedestrian Counts

• Challenge: The YOLOv8 output (shape (1, 84, 8400)) produced overlapping detections, leading to inaccurate pedestrian counts.
• Solution: Implemented Non-Max Suppression (NMS) post-processing, converting center-format coordinates (x, y, w, h) to corner format (x1, y1, x2, y2) for filtering with an IoU threshold.

Future Enhancements

• SQS Integration: Add an SQS queue as a Lambda destination for result buffering.
• Real-Time Streaming: Replace S3 with Amazon Kinesis for live data processing.
• ETL Automation: Use AWS Step Functions to orchestrate the pipeline.
• Optimization: Collaborate with cloud teams to further reduce latency.

License

[MIT License]

---