Trend Report: AI Driven Computer Vision

Companies are showing more and more interest in AI-powered computer vision technology, and an increasing number of technology providers are ready and able to supply this demand. The global computer vision market is building steadily, with researchers predicting growth from $9.40 billion in 2020 to $41.11 billion in 2030, a decade of CAGR (compound annual growth rate) at 16%.

As computer vision proves its worth in specific applications around the world, it already looks set to optimize many sectors. This trend report highlights the use of computer vision in many areas of logistics, from dimensioning and safety to route optimization and demand prediction. At the same time, it can also be applied in other industries such as – retail, healthcare, disaster response and recovery, and manufacturing. This illustrates the enormous potential this AI-based technology has in the supply chain.

The DHL Logistics Trend Radar identifies computer vision as a trend that will become part of the standard way of operating in the logistics industry within the next five years, underpinning and driving future logistics success by enabling more automated and efficient processes as well as sustainable and safe operations.

AI enables computers to “think” and computer vision allows computers to “see and understand.” Computer vision systems gather information from visual inputs like digital images and videos. By collecting and crunching this visual data using algorithms, these systems can then make suggestions and even perform actions.

Since birth, every sighted person has been learning how to tell objects apart, estimate object distance and speed, spot visual anomalies, and interpret what we see. This is the basis of AI-powered computer vision as well.

Computer vision systems, specifically their algorithms, must be trained in the same way, and this is done using visual data. The training process is accelerated by providing vast amounts of digital input. These systems never get tired and can quickly exceed our human capabilities of detecting and reacting to visual inputs. Computer vision accuracy rates for identifying and classifying objects increased from 50% to 99% in less than a decade.  

Computer vision systems learn by looking at vast quantities of high-quality visual data. They repeatedly analyze this data until they recognize images and learn about image differences. How is that done? Using two different technologies:

•  Deep learning, a machine learning approach, employs algorithms to autonomously extract insights from visual data through the utilization of artificial neural networks, continuously enhancing its knowledge from the available information.
• A convolutional neural network (CNN), which breaks images down into tagged labels and performs the math on these labels to repeatedly check prediction accuracy.

Computer vision is likely to soon unlock many more opportunities, thanks to technology advances in depth perception, 3D reconstruction, and interpretation of dark and blurred images. It’s clear that deep learning has moved from the conceptual realm to practical application as many computer vision applications, from facial recognition to self-driving vehicles, make use of it.

Today’s computer vision systems are deployed in various ways.

The most well-known application is image classification. The system sees an image and predicts it belongs to a certain class (e.g., a human, a pair of safety goggles, a forklift).

Another familiar application is object detection, also known as machine vision. The system not only classifies an image but also takes note of (tabulates) its appearance. Once an object has been detected, it can be tracked. Object tracking is often done using sequential images and video feeds.

A further use of computer vision systems is content-based image retrieval, which helps to increase the accuracy when it comes to searching for and retrieving digital images.

Computer vision imagery data is subjected to various processes including image processing (stitching, filtering, pixel counting), image segmentation (partitioning into multiple segments to simplify or change the representation into something that is meaningful and easier to analyze), blob checking (looking for discrete spots of connected pixels as image landmarks; blobs often represent optical targets for observation, robotic capture, or manufacturing checks), and pattern recognition (algorithm-based template matching to find patterns using machine-learning methods).

Computer vision is already proving its worth in a vast range of applications. In the next chapter we explore key solutions that are deployed in retail, healthcare, disaster response and recovery, and manufacturing. With these real-world computer vision examples, it is easier for us to understand the potential of computer vision applications in logistics and the supply chain.