A quick-service restaurant wants to count customers in line at a register and alert managers when the line is too long. Locations have very limited outbound bandwidth and cannot stream multiple videos to the cloud. Which solution best meets the bandwidth constraint and delivers the line-length alerts?

The correct answer leverages AWS DeepLens for edge processing, which is crucial given the limited outbound bandwidth. By training a custom model in SageMaker and deploying it to DeepLens, inference (people counting) happens directly on the device. A local Lambda function on DeepLens then processes these counts and sends a small SNS notification only when the line is too long, minimizing data transfer. Incorrect options: Streaming video to Kinesis Video Streams and then to Rekognition in the cloud would consume significant bandwidth, violating the constraint. While DeepLens can use built-in Rekognition, it might not be optimized for specific people-counting scenarios, and the local Lambda function would still need to process the Rekognition output. Sending images to a SageMaker endpoint in the cloud for inference would also require substantial bandwidth for each image, which is not feasible with the given constraint.