Video Analytics – An Introductory Overview

Overview

Increasing hours of video footage, combined with the limits of human biology, make video analytics software essential to processing large amounts of video streams. Information is most valuable when it is most needed, and is also as valuable as the incidents that can be actively captured and acted upon in real-time. Video management systems assist security and surveillance personnel by monitoring video streams 24/7 and alerting them to activity which requires attention.

Advances in storage capabilities, video and image resolution drove video analytics adoption over the last decade. The global video analytics market was valued at $2.77 billion in 2017 and is estimated to increase to a staggering $8.55 billion by 2023. GPUs make it easier to process videos on low-cost accelerators, making the case for advanced video analytics. Signal stabilizing technologies also improve the effectiveness of video analytics which relies on quality video streams.

Video analytics has use cases across many industries. Some examples include:

◉ Retail – Counting customers in a store, tracking movement, optimizing store design and merchandise restocking

◉ Transportation – Left luggage identification in airports

◉ Healthcare – Thermal imaging for elevated temperature monitoring

◉ Manufacturing/Industrial/Construction – Quality control, and health and safety compliance

◉ Food Processing – Quality control

◉ Entertainment and Sports – People counting to manage crowd traffic

◉ City Operations – License plate recognition and vehicle counting for urban planning

◉ Law Enforcement – Searching video content to help investigations

Multiple video streams are often paired to enable use cases, and combined with IoT sensor data for comprehensive insight and situational awareness. Video management systems can also integrate with third party security apps to help organizations take a holistic approach to their video analytics strategy.

How Video Analytics Works

Incoming streams of video are dense combinations of pictorial and audio data packaged as a collection of consecutive frames. For analytics, an individual frame would not provide any more insightful information than a typical photo. The sequential continuity of consecutive frames provides the dynamic needed to extract insight from video data.

Video data is processed in two stages: first motion detection and Aanalysis, followed by pattern recognition. During motion detection and analysis, changes in pixel content are monitored to identify movement, then pattern recognition classifies objects in motion, their trajectory, and considers other moving objects.

Analytics-enabled cameras use a mathematical function to detect objects in motion by calculating the difference between frames. If the difference is anything but zero, movement is said to have occurred. This is a simple task computationally, while analyzing movement requires more complex computational gymnastics, mathematical functions included. This is where AI comes in. Movement is viewed as a trajectory drawn from an initial frame and tracked to an object’s position in subsequent frames.

As an example, consider tracking one automobile’s movement in traffic for five seconds without confusing its trajectory with other automobiles in the extracted video stream. An image segmentation algorithm segments the vehicle in the initial frame and connects the identified image from frame to frame, thus drawing out the trajectory. Basic Computer Vision solutions with reasonable computational power support this easily with CPU cores and a lavish amount of RAM. In the case of 30 automobiles, 30 segmented images in 30 different spots need to be tracked without confusion when images overlap. For video analytics solutions to be valuable here, they need to process many objects per frame as well as their movement across frames, when the average rate is 60FPS. Imagine looking at 60 photos in one second and fully understanding what was in each photo. More advanced computational capacity is needed to optimize for scale in scenarios like this.

Tools and Development

Deep learning techniques used in intelligent video analytics vary. One common approach is converting video frames to image files and applying Convolutional Neural Networks (CNN) to detect objects in each frame. Hybrid models of CNNs and Recurrent Neural Networks (RNN) are recommend for motion analysis. It’s outside the scope of this blog to cover all candidate tools and frameworks available to implement video analytics applications; however, here are some resources where a variety of options are reviewed for tooling and development:

◉ Open Source Tools – 33 Open Source products on GitHub available for Video Analytics. The major difference amongst these is the use case they enable.

◉ VidSaga, a global video marketers community suggests 2020’s top 10 commercial Video Analytics tools for Business Intelligence.

The core power of such tools and other popular frameworks are the APIs they offer to enrich correspondence in algorithm implementation. Google Object Detection is popular for rapid creation of object detection models. It provides APIs leveraged over 330,000 classes in the  COCO Data Set for object classification. It also allows use of libraries like OpenCV for segmentation and enforces appropriate object-labeling as prerequisites for object tracking. This blog walks through a step by step approach to developing a video analytics application in Python, using the TensorFlow framework and OpenCV library, also applied via Google Object Detection APIs. For a complete implementation example here’s how a soccer game is analyzed. Players (objects) are detected by segmenting them in video frames using OpenCV to assign attributes like jersey color, then labeled and tracked using Google Object Detection APIs.

Source: delltechnologies.com