IntelliCam

Security Camera with Artifical Intelligence processing (Raspberry PI -> AWS Rekognition)
by Matthias Gemelli, Oct 2017

0 - IntelliCam Features

• automatic image recording & upload (Motion triggers s3_upload.py)
• AWS S3 bucket triggers AWS Lambda on each upload (for *.jpg files in upload/ folder)
• AWS Lambda processes each image with AWS Rekognition and:
• reads alarm list (Person, Car, ...) from S3 (reko_alarms.csv)
• calls AWS Rekognition for the uploaded image, compares image labels against alarm list
• when alarm triggered, moves image to folder jpg_alarm/, otherwise to jpg_quiet/
• saves all recorded image labels to folder csv_alarm/, otherwise to csv_quiet/

Planned Improvements & Features:

• analysis of alarms (over time, labels, false negatives...)
• analysis of AWS cost (is this solution affordable?)
• better code structure (config files, s3 folder structure...)
• notification routine (e.g. S3 Lambda triggered by jpg_alarm/ file)
• Image processing (Thumbnails, Daily Gallery)
• email notifications (e.g. AWS SES)

1 - Background & Initial Challenge:

Setting up a Raspberry-PI-powered security camera is quite easy - thanks to software packages like [Motion] https://motion-project.github.io/ or [MotionEyeOS] https://github.com/ccrisan/motioneyeos.

The challenge? After setting up the Raspi camera I recorded 300..400 images. Per Day.
Way too many pictures: my car turning into the driveway, neighbors cats, etc... But I only wanted to see who is entering my property! I finally switched off the camera.

Until I stumbled upon...

2 - The Solution: "AI" with Amazon REKOGNITION

I read the excellent blog posts by Mark West and realized that I could simply filter the images with Amazon's "Artificial Intelligence" service AWS REKOGNITION. Mark explains how he let AWS Rekognition automatically detect (and notify him) if there were people in the picture. Exactly what I was looking for!

• [Mark West: Smarten up your Pi Zero Web Camera with Image Analysis and AWS, pt1]
  https://www.bouvet.no/bouvet-deler/utbrudd/smarten-up-your-pi-zero-web-camera-with-image-analysis-and-amazon-web-services-part-1
• [Mark West: Smarten up your Pi Zero Web Camera with Image Analysis and AWS, pt2]
  https://www.bouvet.no/bouvet-deler/utbrudd/smarten-up-your-pi-zero-web-camera-with-image-analysis-and-amazon-web-services-part-2
• [Mark West: Github Repo] https://github.com/markwest1972/smart-security-camera/

Also stumbled over the post by Amit Agrawal on the AWS blog, providing me with a simple and illustrated tutorial:
[Amit Agrawal: Capture and Analyze Customer Demographic Data Using Rekognition & Athena] https://aws.amazon.com/de/blogs/ai/capture-and-analyze-customer-demographic-data-using-amazon-rekognition-amazon-athena/

3 - My Implementation (RPi --> S3 --> Lambda)

If you want to follow my implementation, I suggest the following steps:

1. Try the AWS Rekognition Demo with your sample images
2. Setup AWS S3, IAM Role and AWS Lambda - test by manually uploading some images. Use this [code for Lambda]https://github.com/MatthiasGemelli/IntelliCam/blob/master/IntelliCam_lambda_py.md
3. Setup RPi with Motion and automatic image upload to S3
4. Image download

3.1 - Try the AWS Rekognition Demo

Ideally you have some sample images from your Raspberry camera, upload them to the AWS Rekognition Demo at https://eu-west-1.console.aws.amazon.com/rekognition/. This should give you an understanding what labels to expect (People, Person, Automobile, Building...).

3.2 - Setup AWS IAM role, AWS Lambda, AWS S3 bucket with trigger

Following Amit Agrawals post you should:

• create a new AWS IAM role with Lambda Execution and full access to S3 and Rekognition
• create a new AWS Lambda routine using my code
• create a new AWS S3 bucket with trigger linked to the new Lambda routine
• upload some image to your S3 upload folder
• check Cloudwatch for log messages and troubleshooting
• check the processed images and labels

3.3 - Setup RPi with Motion and automatic S3 image upload

Using my S3 upload scripts...

3.4

4 - Lessons Learned, Next Steps

My Conclusions:

• Shit. It works!

• Reddit AWS post https://www.reddit.com/r/aws/comments/759ovz/aws_rekognition_fail/

• Stack Overflow post https://stackoverflow.com/questions/46649946/aws-rekognition-fails-to-detect-people-in-4-5