Multimodal Perception Data and One-Stop Algorithm Training for AI Empowerment in Jiangsu Courts

Building a One-Stop Training Platform Shared by All Courts in the Province

To ensure the data security of the court’s dedicated network, Jiangsu courts rely on the built-in one-stop training platform of the Jiangsu Court Perception AI Empowerment Platform, utilizing different model training solutions and data intelligent labeling technology to achieve rapid iteration and optimization of various audio and video algorithms.

Model Training. In response to violations such as unauthorized video recording and physical conflicts in courtroom areas, Jiangsu courts leverage the one-stop training platform to create 12 types of training models, including image single-label classification, image multi-label classification, image multi-attribute classification, and object detection models, conducting targeted training to fully meet the AI application needs of courts in various scenarios and fragmented streaming media.

Intelligent Labeling. Pilot courts in Jiangsu have quickly fine-tuned the foundational models using intelligent labeling technology for audio and video data based on the models trained and generated at the Jiangsu High People’s Court, automatically labeling the incremental audio and video data they collect, completing optimization training through simple result adjustments and confirmations.

Creating an Algorithm Repository Supported by Unified Management and Scheduling of Intelligent Algorithms

To address the inconvenient issues caused by different sources and formats of audio and video across courts, pilot courts in Jiangsu analyze the sources and real-time needs of different audio and video through an algorithm repository, and achieve recognition and access to audio and video algorithms provided by different manufacturers through algorithm scheduling of cloud edge devices like cloud centers and front-end devices, realizing unified management of multi-category and multi-manufacturer audio and video algorithms from aspects like defining algorithm package structure specifications, structured description specifications, and interface specifications.

In addition, pilot courts apply ultra-wideband wireless communication technology (UWB) and the principle of TDOA (Time Difference of Arrival) for spatial positioning, deploying positioning bases in security inspection rooms, litigation service halls, various floors of trial courts, and specialized business rooms. When registering visitors, temporary visitor passes with UWB positioning functions are issued, achieving precise positioning of visitors from the entry security inspection registration to the end of their visit. In the criminal trial guarantee work, by having criminal defendants wear positioning wristbands, functions such as full-process positioning of the defendants, anti-tampering alarms, and escape alarms are realized.

Real-Time Detection of Abnormal Behavior in Court Trials and Petition Scenes and AI Video Labeling

In response to certain parties being improperly dressed, smoking, or eating during remote court sessions, pilot courts in Jiangsu utilize the Jiangsu Court Perception AI Empowerment Platform to automatically identify these abnormal behaviors, labeling the video stream and overlaying event trigger frames and scene management frames to display warning information in real-time, thus promptly correcting violations and enhancing the normativity of court trials and the seriousness of justice.

Real-Time Interactive Audio and Video Across Multiple Business Scenarios

Jiangsu pilot courts leverage the Jiangsu Court Perception AI Empowerment Platform to enhance audio and video interaction quality. In response to network conditions during cross-network court sessions, multiple weak network countermeasures such as anti-network fluctuations, packet loss compensation, adaptive bitrate, adaptive frame rate, adaptive video and image, dynamic multi-streaming, and adaptive sending and receiving buffering are implemented, achieving audio echo cancellation, noise reduction, background noise elimination, and audio-visual synchronization effects. Additionally, based on facial recognition technology, background image overlay and person extraction are achieved, supporting remote trials, remote interrogations, and remote mediation, effectively breaking geographical limitations and improving case handling efficiency.

Online Retrieval of Multi-Modal Electronic Evidence and Key Information Extraction and Localization

The pilot courts in Jiangsu rely on the Jiangsu Court Perception AI Empowerment Platform to achieve key information extraction and quick localization of audio and video evidence. Second-instance judges can extract and automatically label key audio and video segments from first-instance trial data, allowing for rapid retrospection of the first-instance trial process.

At the same time, the platform supports the integration of case audio and video materials from the electronic dossier system’s material transfer module, providing an upload entrance for audio and video evidence via internet litigation service platforms or apps, enabling parties to upload evidence materials independently and supporting electronic material data integration. After the audio and video evidence is integrated into the case, pilot courts conduct playability tests, and for audio and video that cannot be played normally, a unified transcoding process is carried out, generating a playable audio and video file for online retrieval while retaining the original file.