Analysis

Video analysis is a critical application area of artificial intelligence (AI). With the improvement in GPU computing capabilities, various industries are integrating video AI to assist in making traditional tasks more intelligent. Tasks that used to take months to complete or relied on human visual recognition can now be accomplished in milliseconds by providing ample training data for model development. Despite its simplicity in concept, many enterprises struggle with the successful implementation of AI video recognition technology. One of the main reasons is a lack of expertise in peripheral knowledge related to video, resulting in suboptimal recognition performance and potential project setbacks.

Furthermore, system integrators often need to decide whether to use cloud or on-premises architecture for AI video analysis. In a cloud-based architecture, captured video are directly uploaded to cloud services like Google, Microsoft, or Amazon, where cloud servers analyze the video and return results. Cloud services offer advantages such as streamlined hardware requirements and a wide range of AI models. However, they also come with challenges related to image privacy and latency, which can limit certain applications. In specialized video applications, especially those requiring low latency, edge computing becomes increasingly important. Enabling machine vision capabilities can be achieved by adding capture cards or cameras to edge computing devices, allowing real-time AI analysis. This trend is influenced by continuous improvements in hardware technology, particularly the enhancement of GPU computing power for edge computing.

YUAN has over 30 years of experience in video processing and has accumulated extensive industry knowledge across various domains. This extensive practical experience gives us a competitive edge that is difficult to match. Therefore, for enterprises that cannot independently develop AI, YUAN can assist by providing a variety of native video analysis models and rich industry experience through the NexVDO SDK, helping clients achieve software intelligence and seamless integration with their respective industries. For partners already specializing in AI image analysis, they can also benefit from the support of the NexVDO SDK's capture, recording, and streaming modules. Recognition analysis often depends on image quality, and YUAN excels in front-end video processing, achieving efficient resolution and color space conversion, giving clients full control over image quality and capture speed. In addition, YUAN conducts image recording and streaming in post-processing, reducing the burden on clients in terms of R&D.

YUAN's NexVDO SDK divides AI analysis into three main modules: Video Content Analysis, Biometric Recognition, and Behavior Analysis.

Video Content Analysis

Video content analysis aims to understand the context of specific objects or events within an image. The recognition process includes image segmentation and object detection. For instance, in the context of an autonomous driving system, image segmentation distinguishes different objects in an image, such as roads, pedestrians, vehicles, trees, and the sky, while object detection further identifies the specific type of vehicles, such as distinguishing between small cars and large buses. There is no one superior method over the other; the choice depends on the problem at hand to ensure faster and more accurate results.

In NexVDO SDK, identified objects can be further utilized for classification counting, feature matching, and text recognition, providing a user-friendly API for developers. Classification counting finds applications in various industries, such as retail for people and vehicle counting and monitoring high-traffic zones in shopping malls. Feature matching can be applied in medical ultrasound tumor identification, security surveillance video retrieval, cross-camera tracking, factory defect detection, and production automation. Text recognition includes applications like intelligent parking lot license plate recognition and medical information digitization.
 

Biometric Recognition

NexVDO SDK offers rich facial information suitable for various biometric scenarios. The facial recognition in NexVDO SDK achieves an exceptionally high 1:N recognition rate of 99.8 and has consistently ranked among the top performers in the NIST's FRVT competition. The recognition model supports multi-angle recognition, even accurately identifying side-profile faces turned to approximately 120 degrees. Additionally, the SDK covers the extraction of 68 3D facial feature points, facilitating detailed expression recognition, including six emotions: calm, surprise, happiness, anger, disgust, and fear. This can be applied in beauty effects and facial animation special effects.

In the retail industry, we help businesses analyze potential customer preferences by collecting video data from intelligent dashboards, accurately analyzing customer age, gender, and expressions, leading to more precise advertising content delivery and increased product conversion rates. For online educational courses or remote education providers, integrating expression analysis can assist teachers in analyzing classroom performance based on student expressions, quantifying learning effectiveness. NexVDO SDK provides a user-friendly API interface in the field of facial recognition, covering a wide range of applications, from general expression recognition to precise facial recognition, making it easy to integrate into software development across various industries.
 

Behavior Analysis

This module is applied in scenarios that require the analysis of continuous images. NexVDO SDK can identify 17 key points from head to toe, enabling the recognition of current behaviors, such as raising hands, standing, sitting, lying down, and crossing. When behavior is analyzed along the timeline of continuous images, it expands to encompass various behavior recognition, such as flipping, falling, wandering, throwing objects, and other types of behavior analysis.

Learning from the human skeleton and timeline, a wide range of applications can be developed. For example, in baseball pitcher training, clients use 1080p240 high-speed image capture cards to analyze each frame, conduct posture analysis using the skeleton key points provided by NexVDO SDK, and guide them to achieve superior performance based on quantified data. In smart retirement centers, the behavior of elderly individuals requires close attention, and immediate notification to caregivers can be achieved through zero-contact identification analysis in the event of a fall, leaving video records when such incidents occur. Additionally, YUAN collaborates with educational institutions to gradually build what is known as an intelligent classroom application. Classroom materials combine with camera footage to simultaneously capture images of student classroom interactions, and through posture analysis, various student behaviors can be analyzed in real time, even when the teacher moves around.
 

Through the application of NexVDO SDK, we can clearly observe that AI video analysis has already brought about revolutionary changes in various industries. YUAN's advantage lies in our full integration of NVIDIA and Intel GPUs. These technological packages effectively enhance the efficiency of deep learning inference and deployment, helping developers shorten their development cycles. From image capture to intelligent analysis, we provide you with significant acceleration!