,allowExpansion)
AI and Computer Vision in Security and Quality Assurance
Augment Human Oversight and Enhance Process Control
Real-time computer vision is creating tangible improvements in industrial and security settings, giving machines the ability to interpret visual data from images and videos, and allowing for the instantaneous analysis of events as they unfold
The essence of Real-Time Processing
At its core, real-time computer vision is defined by the ability of a system to process visual data and provide an output almost instantaneously. This capability is the key to creating solutions that can intervene in and improve dynamic processes, whether for ensuring security or managing production lines. Achieving this requires not only clever algorithms but also a hardware and software stack optimised for performance.
High-speed inference with NVIDIA TensorRT
A critical component of the software stack is NVIDIA TensorRT, a high-performance library designed specifically to accelerate the inference speed of AI models. A key technique it employs is precision reduction, where the calculations within the neural network are performed using lower-precision numbers. This significantly speeds up computation while being carefully managed to avoid any loss of accuracy in the final result. TensorRT is compatible with popular deep learning frameworks, including the PyTorch library used in these solutions, and it is instrumental in maximising the performance of the underlying GPU hardware.
Powerful and private Edge deployment
The hardware foundation for these real-time applications is the NVIDIA Jetson AGX Orin. This is a powerful yet compact edge device designed for high-performance AI applications. Processing data "on the edge"—meaning locally on the device itself rather than sending it to a remote server—offers several key advantages. It significantly reduces latency, as data does not need to travel over a network; it saves bandwidth; and it enhances privacy and security, as sensitive visual data remains on-site.
Challenges of Airport Security
Security within an airport environment is a matter of paramount importance, involving numerous checks and a large contingent of staff to ensure the safety of thousands of passengers daily. Technological innovation plays a crucial role in supporting and improving these rigorous procedures. However, maintaining constant, unwavering alertness is a significant challenge for human personnel. It is here that artificial intelligence can provide a vital layer of support, augmenting human capabilities to improve security checks, such as those performed during the scanning of boarding passes. Computer vision technologies are playing a critical role in addressing security vulnerabilities in these airport environments, particularly the issue of tailgating and perimeter violations.
Fighting tailgating
and climbing detection
Tailgating occurs when an unauthorised individual closely follows a legitimate passenger through a gate before it closes, bypassing identity checks. To counter this, a real-time computer vision system has been developed that continuously tracks individuals as they approach and pass through e-gates. By combining motion tracking with gate status data and temporal analysis, the system can accurately determine whether each entry is legitimate. The open-source model RTMDet is used for efficient and accurate person detection, while a tracking module assigns unique identifiers to individuals across video frames, enabling continuous movement analysis. If two people are detected passing through during a single gate cycle or within an implausibly short time interval, the system classifies it as a tailgating event and immediately triggers an audible alarm and user interface alert for security personnel.
Another critical use case is about individuals attempting to jump over the queue balustrade to bypass the boarding pass scan. This requires a more detailed understanding of human posture and movement, which is achieved using the RTMPose model. The system first detects individuals and then analyses 17 key body points—focusing particularly on the hips—to determine whether someone is attempting to climb the barrier. Based on a person’s proximity to the balustrade, the system activates pose estimation in real time and evaluates their motion. If climbing behaviour is detected, an alarm is triggered to alert staff. By combining person detection, tracking, and pose estimation within a unified computer vision pipeline, the system significantly enhances situational awareness and supports human oversight in detecting breaches that would be difficult to spot through manual monitoring alone.
The importance of low false positives
A crucial metric for the success of any automated security system in a busy airport is its false positive rate. Thousands of passengers, often with luggage or interacting with each other, pass through these areas daily, creating complex scenes that could potentially trigger false alarms. An excessive number of false alarms would be a distraction rather than an aid to security personnel, rendering the system unusable. The developed systems boast a false positive rate of under 0.2%. This is a vast improvement on previous sensor-based systems for tailgating, for instance, which had a false positive rate exceeding 5%, a figure that could translate into more than 600 false alarms per day.
Leveraging AI and 3D point Clouds for Industrial Quality Control
In industrial environments where precision is critical, AI is increasingly being used to support quality control by offering consistent, real-time monitoring that overcomes the limitations of human inspection, such as fatigue or subjectivity. While traditional 2D imaging can detect surface-level defects, some applications require a more detailed understanding of an object’s shape and structure. To address this, AI systems can analyse point clouds—high-resolution 3D representations made up of thousands of spatial data points. However, processing such dense data in real time poses challenges, as there are no standard compression methods for depth information, and computational demands are high. Efficient techniques like cropping to isolate relevant regions and downsampling to reduce data size are therefore essential for making real-time 3D inspection both accurate and practical.
AI-Driven 3D inspection for pallet assembly monitoring
This system employs advanced AI and 3D imaging to monitor a robotic arm as it assembles pallets by layering boxes. Using a Time-of-Flight (ToF) 3D camera, which combines RGB and depth streams to generate precise point clouds, the system inspects each completed layer for quality and accuracy. Two key performance indicators (KPIs) are assessed: the fill rate of the pallet layer and its alignment. The camera’s precision allows detection of misalignments as small as two centimetres from a distance of six meters. The inspection pipeline begins by capturing static scenes after the robotic arm has moved, transforming the point cloud’s orientation to the world’s coordinate system, and selecting optimal frames through density analysis—all performed with multiprocessing for real-time operation.
Following frame selection, the system cleans and optimises the point cloud by removing noise caused by metallic reflections through clustering algorithms, focusing on the largest connected component. It handles empty pallets to set baseline KPIs and uses the fast MobileSAM model for segmenting the pallet. Cropping and downsampling further reduce data complexity. A hybrid 2D/3D approach then accurately estimates the contours of the pallet layer. Finally, the KPIs are computed by comparing the polygonal area of the layer against the empty pallet and checking for alignment anomalies when points fall outside defined boundaries, ensuring high-quality pallet assembly.
Demonstrating tangible results
The effectiveness of this system in industrial contexts is evident in its results. Over a three-week trial period, the number of defects detected by the system was greater than the average number of annual complaints the client received for faulty pallets, suggesting that many defects were previously going unnoticed. The performance metrics were outstanding: anomalies were detected in only 0.009% of the analysed layers, highlighting their extreme rarity, with an exceptionally low false positive rate of 0.001%.
Technology Reply is the Reply Group company specialized in designing and creating innovative solutions based on Oracle technologies. With deep expertise and experience gained over more than 25 years of activity, Technology Reply supports its clients in innovation and digital transformation processes, with a team of professionals capable of intervening at all stages of a project lifecycle: analysis, design, implementation, launch, and operation. Technology Reply positions itself as a reference partner for Oracle on-premise and cloud technologies: BigData, Analytics, eCommerce, Mobile, Identity and Access Management, and Content Management. With its Agile culture, Technology Reply is becoming a leader in emerging technologies, implementing Artificial Intelligence, Machine Learning, DevOps-based microservices, and cognitive technologies, for a wide variety of reasons, sectors, and clients.