Recently, in the MOT Challenge (Multiple Object Tracking) competition, the most authoritative evaluation platform in the field of international multi target tracking, Minivision broke through the monopoly position of internationally renowned institutions such as Google, Microsoft, Facebook, Amazon, NEC, MIT, and proposed the MiniTrack tracking scheme, ranking first in the international rankings of 8 evaluation indicators. Among them, the core technical indicators MOTA and HOTA rank first.

Mini Track tracking solution ranks first in multiple indicators

(Bold text indicates the best indicator on the current list)

MOT Challenge was jointly founded by the University of Adelaide, the Federal Institute of Technology of Zurich, and the Technical University of Darmstadt. It is the most authoritative evaluation platform in the field of international multi-objective tracking and can be said to be a "battleground" in the CV industry. In recent years, major AI research institutions and enterprises around the world have engaged in technological competition here, competing for algorithmic hard power. So far, more than 100 institutions have participated in the evaluation.

Obtaining excellent rankings in the MOT international competition poses a high challenge.

On the one hand, the technical strength of the participating teams is at the world's leading level, and the competition is very fierce; On the other hand, multi target tracking requires handling complex lighting, line of sight occlusion, blurring, and other challenging situations. Due to the immaturity of technology, problems such as low accuracy, slow speed, motion blur, and ID jumping are common and pose extremely high challenges.

Anchor multi target tracking to see how Xiaoshi can overcome difficulties

Minivision continues to use the Tracking by detection tracking framework to improve target detection, data association, and time-domain association. It proposes a smooth association method, which achieved excellent results ranking first in the MOT2017 competition.

racking-by-detection pipeline

In multi target tracking, in general, external factors such as target occlusion, small targets, and blurring can lead to a decrease in detector performance. Smooth association can effectively address such issues and improve the stability of multi target tracking by comprehensively considering changes in tracking targets and changes in the output of target detectors.

Smooth-association serializes and analyzes the tracking target. Based on the stability of the detector output, multi-level dynamic matching is used for similarity matching and data association to more effectively reduce the impact of detector instability on the tracker.

Single threshold match & dynamic threshold match in tracking procedure

In the large-scale production application process of computer vision technology, especially in the field of fine-grained video structuring, multi-objective tracking is a fundamental analysis task with high technical requirements. As a necessary link, this technology runs through various video analysis tasks. On the one hand, tracking technology can improve target recall rate and compensate for detection accuracy in sequence analysis; On the other hand, the tracking algorithm provides a consistent label of scene targets in the time series, enabling the algorithm scheme to perform three-dimensional analysis of targets in the temporal and spatial dimensions. In addition, as an algorithm unit for real-time analysis, multi target tracking also requires extreme performance.

The key to Minivision's victory: algorithm implementation, scenario validation, and deep learning

Since the 18th National Congress, Chinese government has attached great importance to the innovation and tackling of key core technologies. The Outline of the 14th Five Year Plan states: "Looking ahead to 2035, China will achieve significant breakthroughs in key core technologies and enter the forefront of innovative countries." In March 2022, Li Keqiang pointed out in his "Government Work Report" to the National People's Congress that it is necessary to deeply implement the innovation driven development strategy and continue to promote key core technology research.

Without going through countless trials and tribulations, one cannot attack key core technologies. Minivision focuses on technological innovation and has accumulated over 200 self-developed visual algorithms in recent years, obtaining nearly 90 patents for visual related algorithms. Minivision's self-developed algorithm covers numerous tasks such as traffic statistics and abnormal behavior recognition under intelligent video monitoring, providing basic data for subsequent event warning and intelligent decision-making in business. The relevant technology has been implemented in many scenarios such as smart communities, smart campuses, and smart parks, widely covering major provinces and cities across the country.

The sharpshooter is fed by bullets, "Minivision's original algorithm is honed in countless scene applications.

In smart communities in Jiangsu, Chongqing, Liaoning, and other places, Xiaoshi empowers intelligent perception hardware with tracking algorithms, connects front-end devices such as SDC cameras, vehicle road gates, pedestrian gates, and access control, and achieves functions such as high-altitude throwing recognition, personnel gathering recognition, vehicle detection, fire occupation, and dog walking without rope recognition. Through the data center, Xiaoshi completes data collection, analysis, processing, and presentation, assisting the community in moving from civil defense to intelligent technical defense.

Minivision Building Smart Communities in Chongqing

Minivision Building Smart Technology Prevention Community in Jiangsu Province

Minivision has assisted more than 300 schools in Jiangsu, Zhejiang, Henan, Guangdong, Shanxi, Liaoning and other provinces and cities in achieving intelligent upgrades. The system can achieve various functions such as high-altitude throwing detection, student running and fighting behavior recognition, crowd gathering recognition, and danger zone intrusion warning, helping campus security.

Minivision Building a Smart Campus for Hangzhou Primary School

In the smart park project of Shenyang Hunnan District Government, Minivision achieves six major functions: intelligent attendance, intelligent epidemic prevention, intelligent security, intelligent dining, and intelligent conference check-in. In addition, by empowering thousands of cameras in the entire area with AI, video surveillance in Hunnan District can capture and identify phenomena such as road occupation, illegal parking of shared bicycles, and illegal parking of motor vehicles in real-time, collect data, and intelligently analyze them, achieving intelligent urban governance.

Minivision Building a Smart Park in Shenyang

Since its establishment, Minivision has been committed to empowering thousands of businesses with AI technology. The top spot in the MOT Challenge this time is a full affirmation of Minivision's multi target tracking technology, and also confirms Minivision's strong algorithm strength in the field of computer vision.

In the future, Minivision will also deepen the application of multi target tracking technology in scenarios, helping various industries to upgrade their intelligence.