On August 26th, the 5th Jiangsu Artificial Intelligence Conference (JSAI2022), hosted by the Jiangsu Artificial Intelligence Society and Jiangnan University, was grandly held in Wuxi.
As one of the conferences with the highest specifications, largest scale, and strongest influence in the field of artificial intelligence in Jiangsu Province, this conference gathers innovative resources in artificial intelligence at home and abroad, showcasing the academic achievements and technological achievements in the field of artificial intelligence in Jiangsu. Leading figures from various industries, universities, and research institutes in the field of AI engage in exchanges and sharing around AI theory, technology, industry, and applications, exploring the future development of artificial intelligence.
At the sub forum event, Hu Jianguo, Vice President of Minivision Technology and Dean of AI Research Institute, brought a theme sharing on "Exploration of Fine Granular Structured Analysis Technology for Large Capacity Video". President Hu elaborated on the application practice of spatiotemporal information fusion in video structured technology, and further analyzed the new possibilities brought by the combination of Minivision Technology MOT international competition champion scheme and industrial applications.
Practice of "Minivision": Algorithm Application of Video Structuring
Video structured technology can convert a large number of original surveillance video images into structured information, completing information mining and data extraction, which is crucial for the intelligent transformation of security.
Xiaoshi Technology utilizes video structuring to capture key information in video frames (such as position coordinates, physical features, motion trends, trajectory information, etc.) through spatiotemporal dimensions, in order to obtain video prediction results and apply them to target tracking. For example, monitoring the wearing and taking off process of protective clothing, identifying fights, detecting high-altitude throwing, and detecting campus running.
The real environment of these four types of scenes is extremely complex, making tracking and detection extremely difficult. The process of wearing and taking off protective clothing includes multiple components such as behavior recognition, monitoring and tracking technology, refined label classification technology, and process management, which require high real-time performance; Fighting detection needs to address the impact of factors such as intense movement, high occlusion, and short duration on recognition accuracy.
The difficulty of high-altitude parabolic detection lies in the complex outdoor environment, small target volume, multiple styles, fast falling speed, and high performance requirements for the detector; Campus running detection faces the challenges of fast target movement speed and complex scenes with a large number of people, requiring matching with high tracker performance.
New possibilities for industrial application of MOT champion solution: MOT17 and MOT-HT21 rank first
The complex real scene places higher demands on detection and tracking algorithms. In the MOT Challenge (Multiple Object Tracking) competition, the most authoritative evaluation platform in the field of international multi object tracking, MiniVision Technology's MiniTrack tracking scheme performed exceptionally well, achieving first place in MOT17 and MOT-HT21 rankings.
For the MOT-HT21 task, the method proposed by Minivision Technology ranks first in the world in 9 evaluation indicators, including MOTA 80.9, IDEucl 73.7, and HOTA 52.7, and is significantly ahead of other methods. Some of the methods used in the competition have also generated gains in practical production projects, especially in scenes such as small targets and camera shaking, which have shown good performance.
The MiniTrack tracking scheme follows the Tracking by detection tracking framework and improves the target detector and tracker respectively.
To improve the detection performance of small targets by detectors, different scale feature layers are fused to enhance the recall rate of small targets; Ignoring mutually occluded annotated samples during the training process improves the classifier's discrimination of targets. The optimization scheme of the tracker adopts a spatiotemporal motion estimation model, achieving the effect of parameter self-learning, and the prediction results are closer to manual annotation. Adaptive prediction tracking box, considering more useful information, greatly improving speed. The entire tracking process is simple, with full matrix calculation to accelerate tracking calculation efficiency. At the same time, it also effectively solves the problem of current methods being not robust to camera shake and non-uniform target motion.
Productization implementation: technical upgrade, performance optimization
The implementation of product-based applications is a process of fully unleashing technological value.
Minivision Technology's high-altitude parabolic camera has achieved a detection and tracking algorithm that takes only 53ms throughout the entire process, achieving real-time analysis, preventing missed detections, and improving the recall rate by 30%; Minimum detection target reached 5x5@1080p The comprehensive detector recall rate in complex environments reaches 98.5%; The ability to track multiple targets has been improved, and the ability to filter abnormal trajectories in parallel has been improved by 25%.
The technological upgrade of video structuring and tracking detection algorithms provides strong support for the continuous optimization of Minivision technology product performance. In addition to high-altitude parabolic cameras, related technologies are also applied to products such as small view astrology edge analysis hosts and MG800 AI comprehensive capability components, which can be applied to serve the entire scene ecology under smart city governance.
In the future, Minivision Technology will also focus on technology research and development to promote the deep integration and development of artificial intelligence foundation layer, technology layer, and application layer.
