Building a Computer Vision Pipeline with RTSP Stream

00:00 Introduction

Emily from ROFL introduces a tutorial on building a computer vision pipeline using an RTSP stream, focusing on data collection, labeling, and deploying a custom model for stateful analytics, such as counting objects and tracking their time in specific zones.

02:00 Setting Up the RTSP Stream

Emily discusses setting up a camera, specifically a Tapo outdoor camera, to generate an RTSP stream URL. She explains how to access camera settings to create a unique stream URL.

06:30 Data Collection Workflow

The tutorial moves into creating a data collection workflow. Emily demonstrates using pre-trained models like Coco and Yellow World to identify objects in images captured by the camera.

12:00 Filtering Detected Objects

Emily shows how to filter out unwanted classes from the detected objects and tweak the confidence levels to improve accuracy.

18:00 Uploading Annotated Data

Once the desired classes are identified, she explains how to upload annotated data back to RoboFlow, controlling the amount of data collected over time.

22:00 Training the Model

After collecting a sufficient dataset, Emily annotates the images and prepares them for training, applying augmentation techniques to improve model performance.

30:00 Evaluating Model Performance

After training the model, Emily evaluates its performance, discussing the identification rates and areas for improvement based on previous mislabels and missed detections.

38:00 Building the Stateful Workflow

Emily explains how to build a stateful workflow that counts objects crossing a line and tracks how long they spend in specific zones, incorporating visualizations for better understanding.

48:00 Summary and Next Steps

The video concludes with a recap of the pipeline created, detailing the vehicle detection, counting, and time tracking functionalities, with a teaser for part two that will cover data export and notifications.