• Helm Chart Version: 1.7.1
• Alembic Version: 5e1303b49061

Features

• Scenario detection using multiple images: The function has been expanded to perform scenario detection based on multiple images. Compared to single frame-based detection, it can reflect more diverse situations, resulting in improved overall detection accuracy.
• Added VM Only detection process feature: Added a VM Only detection process that generates detection notifications with only the VM without going through the VLM. This allows for a lightweight detection flow and allows for faster notification processing in certain scenarios.
• Scenario Importance Setting Function: A function has been added to define the importance as text for scenarios requiring emergency response. Detection notifications that occur in scenarios set to importance are marked as urgent notifications so that you can be recognized with priority.
• False positive/true positive feedback function expansion and chat window display: In addition to the existing false positive feedback, the function has been expanded to allow feedback for true positive positives as well. The entered feedback can be checked in the chat window in the form of false or true positives, and the information is updated when you refresh or re-enter the screen, not in real time.
• Main camera list column configuration change: The model name and target information columns were removed from the main screen camera list, and new columns for checking notification severity, recent notification summary, scenario summary, and webhook settings were added to improve readability from an operational perspective.
• Vision Model information multilingual support: The name and description of the Vision Model have been improved to automatically display in the language depending on the language setting (Korean/English).
• GUI improvements:
  • The color of the Annotation Box has been changed to gray tone.
  • The danger area rectangle draw method has been improved, and the related logic has been moved from Backend to Frontend to improve UI responsiveness and overall performance.

Performance Enhancements

• License behavior performance improvements: Improved communication with the server to occur only when necessary, eliminating unnecessary server synchronization and redundant code execution. We've also optimized the overall responsiveness of the web interface by reorganizing the API interface.
• Improved Notification List query performance: By applying the Lazy Loading technique, we have improved the delay that occurs when users click the Notification icon.
• Backend DB query performance optimization: By applying the Device Data Caching technique, we have improved the use of cache data in memory instead of the existing method of querying the DB every time Camera Frame is processed. This reduced DB I/O load and improved overall processing performance.
• Frontend streaming processing performance optimization: Canvas and Bitmap-based rendering utilizing browser GPU acceleration was applied, and stream processing and rendering were separated into background threads through Web Worker and OffScreenCanvas (Chrome, Edge). This reduces main thread load and improves streaming UI responsiveness. Additionally, the frame processing overhead was reduced by moving the rendering of the dangerous area from the backend to the frontend.
• Frontend LLM chat processing performance optimization: By applying Web Worker, we have improved the processing of WebSocket-based messages (including images) in the background. We also introduced virtualization components to minimize the number of chat DOM renderings, optimizing memory usage and scrolling performance.

Bug Fixes

• Modified webhook validity check logic: We have fixed a bug that required a validation check for the newly registered webhook when registering an additional webhook.
• Fixed a problem with a delay in the occurrence of the first Alert after running the app: Fixed an issue that caused a delay due to the process of downloading the InsightFace model for face mosaic processing when the first Alert occurred after running the app. By pre-building and including the InsightFace model in the app Docker image, we eliminated the delay in initial alert generation.
• Fixed the issue of inconsistent application of object detection sensitivity to the same target: Fixed the issue so that the same object detection sensitivity is consistently applied to the target even when the same target is registered at the same time in the common scenario and the custom scenario.