alwaysAI Reference Applications

• Starter Apps

• Hailo Apps

• NVIDIA Apps

• Camera Apps

• Streaming Apps

• Example Apps

Starter Apps

App	Description
Detector Tracker	Use Object Detection and Object Tracking to follow unique objects as they move across the frame.
Face Detector	Use Object Detection to detect human faces in a real-time camera stream.
Image Classifier	Use Image Classification to classify a batch of images. The classification labels can be changed by selecting different models.
Instance Segmentation	Use Instance Segmentation to detect, segment and classify individual objects in a real-time camera stream or a video file stream. The types of objects detected can be changed by selecting different models.
Object Detector	Use Object Detection to detect objects in a real-time camera stream. The types of objects detected can be changed by selecting different models.
Pose Estimator	Use Pose Estimation to determine human poses in realtime. Human Pose returns a list of key points indicating joints that can be used for applications such as activity recognition and augmented reality.
Person ReIdentification	Use Re-Identification to re-identify person from one stream into another stream.
Semantic Segmentation VOC	Perform Semantic Segmentation on a series of images using a model trained on the Pascal VOC dataset.

Hailo Apps

App	Description
Hailo Real-Time Object Detector	This alwaysAI app performs object detection in real-time on Hailo accelerator/Co-processor.

NVIDIA Apps

App	Description
CUDA Applications	This alwaysAI application set use a CUDA (Compute Unified Device Architecture) interface which allows CNN models to execute on NVIDIA GPUs found on Jetson devices. CUDA APIs give applications executing on the CPU direct access to NVIDIA’s GPU’s virtual instruction set and parallel computational elements.
NVIDIA Real-Time Object Detector	This alwaysAI app performs object detection in real-time on NVIDIA Jetson devices.
NVIDIA Semantic Segmentation	This alwaysAI app performs semantic segmentation on a video clip of a driving scene on NVIDIA Jetson devices.
Object Detector TensorRT batch mode	Use Object Detection to detect objects in videos in batch mode. The types of objects detected can be changed by selecting different models.
TensorRT Applications	This alwaysAI applications set uses TensorRT binaries to do the local inferencing on a NVIDIA Jetson device, these binaries can be found in the alwaysAI model catalog. The model will start with TRT and end with the Jetson device name it should be run on, for example nano. These binaries are the most efficient way to do inferencing on NVIDIA Jetson device. Currently alwaysAI supports TensorRT binaries for Jetson Nano, Xavier NX, and AGX Xavier.

Camera Apps

App	Description
RealSense Object Detector	Use Object Detection to detect objects and the RealSense camera to get the distances in meters to those objects in real-time. The types of objects detected can be changed by selecting different models.

Streaming Apps

App	Description
Dash Interactive Streamer	This application provides a base for an interactive streamer using Plotly Dash (https://dash.plotly.com). This app includes a video stream, which depicts object detection results, as well as a table that updates every 5 seconds with object detection results. You can create more complex applications by adding components.
Interactive Image Capture App	This alwaysAI app builds off of the app found here to provide a customized data collection server that allows the user to take snap shots or record videos and store them for later use.
Multi-Stream Framework	This application demonstrates using the multi-stream framework to process multiple camera streams using multi-processing.
Multiple Camera Streams	This alwaysAI application demonstrates how to incorporate video streams from multiple cameras into a single app.
Simple Video Recorder	This alwaysAI app records and saves video from a USB webcam or CSI ribbon camera.
Video Streamer	This alwaysAI app performs realtime object detection and streams video and text data to a Flask-SocketIO server that can be located on another device.
Video Streamer Using Multiple Containers	This alwaysAI app performs realtime object detection and streams video and text data to a Flask-SocketIO server running in another container managed by Docker Compose.
ZMQ Video Streamer	This repo has two parts: an alwaysAI computer vision app which performs realtime object detection and streams video via ZeroMQ to a Flask server, and the server to display the video stream in a browser.

Example Apps

App	Description
Age and Gender Classifier	Use two image classifiers to classify a batch of human face images by age and gender.
Twilio Computer Vision App	A simple computer vision app that can detect a chair, and which sends a text message letting the user know when a chair has been detected.
April Tag Detector	This repo contains a basic application that shows you how to detect april tags using the AprilTagDetector in edgeIQ.
Barcode Detector	Use Barcode Detection to detect and decode Barcode(s) in a real-time camera stream.
Capacity Detector	This application is meant to detect people from a live or recorded video stream and determine if they are entering or exiting a location. That info is then can be sent via a POST call to a URL endpoint for processing.
Classifier based Image Sorter	This app can use an image classifier to detect any number of target labels, then sort those images into 1 of 2 folders.
Contraband Detector App	This app uses object detection and employs a correlation tracker to monitor a frame for items deemed 'contraband'. For working or learning at home, these items may be headphones, cell phones, books, etc. We use two object detection models in order to maximize the number of objects we can detect given a set of freely available detection models. You can add or remove models and alter labels to suit your needs!
Crosswalk Object Detection	This app performs object detection on a video of a crosswalk in simulated real-time. The app will automatically perform inference on the CUDA GPU on Jetson devices.
Darts	Play darts with alwaysAI and the Intel RealSense depth camera!
Distance Detector	This app will display the distance between two hands in inches. This app uses a web camera as the video stream and uses reference object metrics to approximate distance between two objects in an image.
Face Counter	Use Object Detection and Correlation Tracking to count unique human faces in the frame on a real-time camera stream.
Face Swapper	Swaps the faces of people in a video stream when more than one person is in the frame.
Facial Landmarks with DLIB	Using DLIB to do facial detection and find landmarks.
Gesture Controlled Home Interaction App	This app is designed to let you use your own custom gesture model to activate voice-activated technology, such as Alexa, Google Home, or Siri, using computer vision. After training your own gesture-detection model, using hand signals of your choice, you can use these signals to play audio files that give commands to smart tech.
Hacky Hour 08/06/2020	Source code for the August 6 2020 Hacky Hour! Demonstrates how a computer vision application can be used in manufacturing to confirm that a worker has chosen the correct component for assembly of a product. The CV application finds the roi associated with an illuminated LED on a tray shelve. It then checks to see if the workers hand has picked up the correct component using a hand detection model and overlapping rectangles algorithm between roi and hand detection bounding box. The bounding box of hand detector turns green when the correct component is chosen and remains red until that occurs.
Jumping Jacks Counter	This app counts jumping jacks for a single person on a real-time video stream or a video file. The app defaults to the TensorRT engine on NVIDIA Jetson devices, and the DNN engine elsewhere.
Kalman Tracking Hacky Hour 05/12/2021	Source for the May 12th Hacky Hour! This app demonstrates how to use the Kalman Tracker to track objects in motion.
License Plate Tracker App	This app uses object detection and employs a centroid tracker to monitor a frame for license plates and vehicles. Please note that the model used in this app does not detect the content or read the characters from a license plate. Pair this app with an OCR library to do so.
License Plate OCR App	This app uses object detection and employs a centroid tracker to monitor a frame for license plates and vehicles and then uses an OCR library to read the license plates. This app will work on development machines, such as MacOS and Windows, but not on ARM 64 architectures due to the dependency on PyTorch. Note that we have provided the model file that EasyOCR provides at runtime to avoid model download issues. We have only provided the necessary files for the English option. If you would like to explore other uses of the EasyOCR library, please see the source: https://github.com/JaidedAI/EasyOCR/tree/393c2d966aa37 160bcfdef5f55ba50daff994409. There is a link to the JaidedAI Model Hub on the repository's README.
Classifier based Image Sorter	This app can use any number of image classifiers to detect any number of target labels, then sort those images into 1 of 2 folders if the target label is detected.
Multiple Object Detectors Sample App	This app utilizes two object detection models, with the option of adding additional detection models. This may be helpful for including models that comprise very different libraries. In this example, one model detects numerous objects of small to medium size, and the other has a more limited library but detects some larger objects, such as airplanes, trains, and sofas. As there is some overlap between models, for instance they both detect people, you can compare the prediction confidences. Additionally, the output for each model appears in a separate video frame.
Object Detection Server/Client	Server app sets up an alwaysAI model to run on a device where images can be sent and inferenced. The server then returns the marked-up image to the client, which is stored in the inferenced folder.
Object Detection & Classifier Sample App	This app enables two alwaysAI models to be used simultaneously: one to detect facial objects, and a second to classify the detected faces in terms of age ranges.
Package Detection System App	This app is designed to let you use your own custom package detection model to detect when packages arrive and when they are removed You'll need an alwaysAI account and to have alwayAI installed:
Pedestrian Segmentation	This app expands on the semantic_segmentation_cityscape starter app to build an app that segments pedestrians and bicyclists in each frame of a video, and performs actions based on the results. The full tutorial can be found on the alwaysAI blog.
People Counter App	This app expands on the alwaysAI face counter starter app to both count the number of non-unique people (same person will be counted multiple times if they exit and re-enter the frame) detected and to display basic time metrics on those appearances.
Classify Poses with Pose Estimation and Support Vector Machines (SVM)	This repo contains an alwaysAI app and Jupyter Notebook to classify poses using Pose Estimation from alwaysAI and machine learning tools from Scikit-Learn. This tutorial will walk through training an SVM to classify a few common yoga poses.
Posture Corrector with Pose Estimation Example App	This app uses pose estimation to help users correct their posture by alerting them with a sound when they are slouching, leaning, or tilting their head down, as well as a printed message with specific suggestions for correcting posture. A scale variable is used to adjust the keypoints measurements for different individuals, accounting for greater or smaller natural distances between keypoints used to detect poor posture.
QR Code Detector	Use QR Code Detection to detect and decode QR Code(s) in a real-time camera stream.
Object Detector and RTSP Streamer	Use Object Detection and RTSP Server to stream Object Detection Inference through RTSP stream. The types of objects detected can be changed by selecting different models.
Semantic Segmentation Cityscape	Use Semantic Segmentation to determine a class for each pixel of an image. The classes of objects detected can be changed by selecting different models. This particular starter application uses a model trained on the Cityscapes Dataset. The Cityscapes Dataset focuses on semantic understanding of urban street scenes, and is a favorite dataset for building autonomous driving machine learning models.
Simple Object Counter	Use Object Detection to count the number of each type of object in the frame of a real-time video stream.
Virtual Green Screen with Edge Smoothing Example App	This app expands on an app that uses semantic segmentation to segment out a person from background noise in a video stream and replace the background with an image or blur it out. This app demonstrates how to subsequently smooth segmentation boundaries. This app also demonstrates how to separate your app configuration information into a separate JSON file. For more details on this aspect of the app, please see the original blog post.
Snapshot Security Camera App	This app expands on the realtime_object_detector starter app to build a simple security camera app that takes a picture of each new person who enters the frame. The full tutorial can be found on the alwaysAI blog.
Spatial AI Webinar Apps	Spatial AI is the blending of machine learning inferencing with geometric data from sensors, enabling robots, drones or autonomous vehicles to better understand the world around them.
Temperature Tracker App	This app demonstrates how to use a temperature tracking utility class for the Raspberry Pi.
Thermal Imaging	This repo contains applications that demonstrate how to use FLiR Lepton thermal camera. The FLiR sensor uses microbolometer array that changes resistance as its heated up. By measuring resistance the sensor can determine the temperature of detected objects. The sensor's firmware creates a colored image that encodes that resistance data into heat map. The image can be processed using the same techniques as visible light.
Vaccination Center Assistant	A series of applications to help monitor a waiting room, log vaccination events, and monitor a post-vaccination room, and write out event logs to help inform vaccination center logistics.
Virtual Green Screen Example App	This app uses semantic segmentation to segment out a person from background noise in a video stream and replace the background with an image or blur it out. This app builds off of a methodology for segmenting out areas of interested, which can be found here. This app also demonstrates how to separate your app configuration information into a separate JSON file. For more details on this aspect of the app, please see the original blog.
YMCA APP	In this example we'll be using a simple set of cases to determine if someone is doing a Y, M, C, or A in our image then displaying the letter on the screen if they do creating a fun virtual experience.
YOLOX Object Detector	Use Object Detection to detect objects in a real-time camera stream. The types of objects detected can be changed by selecting different models.