AI-BOX目標識別與警報盒子功能概述
-
OVERVIEW 概述
- 支持基於國外NVIDIA CUDA 框架或國內HUAWEI華為ASCEND框架。
- 全球頂尖識別算法技術,效率非常高,仅需一次识别即可完成所有对象的分类识别。
- 可以应用到雲端服务器中,也可以应用於小型邊緣計算硬件中,部署Windows/Ubuntu,支持CUDA顯卡以及華為升騰顯卡等,可選不同檔次的硬件。
-
CCTV安全與監控领域研究與積累
基于多年的音视频編解碼開發和全球頂尖識別算法的開發時間,AI应用取得更多積累的技術成功,但随着技术的成熟,这些应用也变得越来越多,结合实施商業需求的场景開發了相當多的商業案例,例如
- 非法闖入:夜間無人值守或重要地點進出監控。
- 摔倒偵測:關懷老人,老人跌倒偵測或打架鬥毆跌倒偵測,可用於社區、監獄、廣場、學校等場所。
- 明火辨識:嚴禁煙火的第三方辨識火焰,防止火災,可用於大廈、森林、工場、化工產、卡拉OK等場所。
- 煙霧排放:同明火場景。
- 越界偵測:越過指定邊界,產生警報,一般用於行人辨識與闖入偵測,結合視訊區域偵測。 可用於無人值守場景。
- 工作瞌睡檢測:工作期間睡覺檢測,避免安全事故發生。
- 離崗檢測:工作期間離開崗位偵測,避免安全事故發生。
- 公眾場地聚眾:偵測人員聚集,避免聚眾群毆、人群密集避免踩踏事件發生。
- 攀高偵測:偵測人員是否進行攀爬,避免安全事故發生。
- 人臉抓拍:人臉資料抓拍,可推送給人臉辨識服務進行1&1辨識及陌生人辨識。
- 遮擋偵測:偵測攝影機是否被人為遮蔽或損壞。 協調維運人員進行維修。
- 垃圾滿溢:偵測垃圾桶垃圾是否滿溢,協助環衛人員智慧調度環衛車輛,節省人力與物力。
- 霸佔公共地方經營:偵測是否有佔用道路非法經營,協助管理部門管理,減輕工作量。
- 安全帽識別:工地安全帽識別,提高工地安全。
- 反光衣識別:工地反光衣識別,防止非工作人員闖入工地區域。
- 電動車進電梯:偵測電動車進出電梯,防止火災發生。
- 口罩檢測:明廚亮灶,檢測食品從業人員是否有戴口罩。
- 蟲害辨識:智慧園林,屋苑綠化,透過AI辨識蟲害,智能指導園丁作業。
- 動物辨識:公眾場所管理方面,對協助保全人員實施動物闖入警報,對郊野公園\泳灘等警報非常有效的管理手段。
- 電梯超員偵測:電梯人員是否超載工作,地鐵升降機,舊式工廠升降機等等比較特殊的實施案例非常有效。
- 海邊釣魚檢測:嚴禁公園釣魚,危險海邊,檢測人員是否有公園釣魚或捕魚行為。
- 海邊游泳檢測:嚴禁游泳海邊檢測人員是否有下水游泳行為。
- 人數統計:人數統計或客流統計,透過AI方式統計繪製市場的客流熱力圖。
- 抽菸檢測:明廚亮灶,檢測從業人員是否有吸菸行為,煙火警報等等。
- 泥頭車辨識:街道泥頭車隨意、掉土的事件。
- 打電話辨識:偵測開車是否有打電話行為。
- 機動車/非機動車識別:機動車和非機動車識別。
- 車流量統計:辨識車輛及統計車流量,統計屋苑車輛出入。
以上是总结的商業需求的AI應用,非常实用的场景,这些算法與應用都是通過我們多年技術研究和實踐成功實現並應用,具有非常可行性,可實施性的。
-
AI-BOX 功能列表
以下是Python版本或C++版本的AI-BOX的功能列表:
Web Front-End List:
- Model Management
-
Support uploading trained model files to the AI box.
Support the addition, modification, update and deletion of models.
Supported model query and retrieval. - Device management
-
Support dynamic addition of real-time analysis equipment (rtsp video stream).
Support dynamic modification, update and deletion of equipment.
Support device query and retrieval.
Support viewing the latest pictures and time of device analysis.
Support device analysis polygon range detection parameter configuration.
Support device perimeter detection parameter configuration. - Scene Management
-
Support adding multiple scenes (models) to one analysis device.
Support dynamic modification, update and deletion of equipment in scenes.
Support advanced query and retrieval. - User Authentication
-
Support user password to log in to the backend management system.
Support user password modification.
Support user password encrypted login.
Support user logout - System Management
-
Support dynamic configuration of box encoding.
Support box data storage days configuration.
Support box concurrent analysis device channel number configuration.
Support recognition result label display configuration.
Support alarm linkage recording configuration.
Support alarm data push third-party platform address configuration. - Other functions
-
Support analysis alarm push and preview.
Support analysis of alarm voice broadcast.
Support night and day theme mode switching.
Back-End Functions List:
- Web Interface Development
-
Provide an http interface based on restful style.
Provide management-related interfaces for models, equipment, scenarios, configurations, etc.
Provide user login, token verification and other interfaces.
Provide file upload, download, and image preview interfaces. - Dynamic Model Management
-
Support one device to add multiple analysis scenario algorithms.
Support the addition, update and deletion of dynamic analysis scene algorithms. - Dynamic Device Management
-
Support dynamic addition, modification, update and deletion of equipment.
Support device dynamic enabling, disabling and video analysis.
Support device polygon and line segment perimeter detection parameter configuration and detection control.
Support device analysis and latest picture storage and update.
Support device analysis video disconnection and reconnection.
Support device connectivity detection and real-time analysis. - Dynamic Scene Management
-
Support one device to add multiple analysis scenario algorithms.
Support the addition, update and deletion of dynamic analysis scene algorithms. - Dynamic System Configuration
-
Support dynamic configuration of box encoding.
Support box data storage days configuration.
Support box concurrent analysis device channel number configuration.
Support recognition result label display configuration.
Support alarm linkage recording configuration.
Support alarm data push third-party platform address configuration. - Alarm Real-Time Broadcast
- Support broadcasting alarms to clients.
- Alarm Temporary Storage
-
Support alarm storage for different scene algorithms.
Support timing statistics and cleaning of alarms. - System Alarm Recording
-
Support device alarm linkage video recording (5 seconds before and after the alarm).
Support device analysis and alarm picture storage.
Support video recording and storage via USB or GPIO linkage. - Alarm Asynchronous Push
-
Support configuration of third-party platform address (http and websocket).
Support real-time uploading of alarm information to third-party platforms (information, pictures, videos). - Docking Interface Protocol
-
Support the same LAN HTTP communication protocol management model, equipment, scenes, alarms, etc.
Support subnet-cloud websocket+http communication protocol to adapt to external network usage needs. - AI early warning app platform
- Third-party application platform based on Java/Vue.js/Python AI box docking (multi-tenant sass cloud platform).
- Work Ledger/Client Balance
-
Ability to view the total number of all warnings and the number of pending warnings in real time (warning categories can be dynamically expanded).
Ability to display pending warning information in real time, including warning time, location, pictures, and videos.
Ability to handle pending early warnings, which can be ignored or turned into personal tasks for processing..
Ability to view warning pictures and play short warning videos.
Ability to display personal to-be-processed task information, including task name, deadline, and release time 6. Able to handle personal pending tasks. - Comprehensive Cloud Atlas
-
Ability to display the point distribution of all analysis video equipment in the form of a map.
Ability to play real-time video from distributed video equipment, record videos, and take screenshots.
Ability to view and analyze the alarm information of the video and process it.
Ability to receive real-time warning information pushed in the background in real time and display it on the map.
Ability to see the online and offline status of video equipment in real time.
Ability to view equipment information and operation and maintenance information on the map.
Ability to save video points to favorites and play favorites videos.
The real-time position of the device can be adjusted by moving it.
Ability to view the personal tasks assigned by the administrator and process them. - Early Warning Management
-
Provide historical tracing and search query of early warnings.
Ability to view the pictures, videos and processing status of the warning.
Early warnings can be ignored and transferred to personal tasks.
All warning information can be exported. - Alarm Real-Time Broadcast
- Support broadcasting alarms to clients.
- Device Management
-
Provide search queries for tasks.
Can view pictures and processing status of tasks.
Tasks can be processed in real time 4. All task information can be exported. - System Configuration
-
Configurable system warning SMS push rules (time, warning type, silent period, warning method, notification objects, etc.).
Can view and retrieve personal warning message notifications. - System settings
-
Can manage the project area tree of the current tenant (personal organizational structure).
Equipment that can manage the system (AI box, video equipment.
Can view the video status structure view (one picture shows the access architecture and device status).
Can view and retrieve the online and offline logs of the device.
Can manage all project information of the current tenant (dynamically switch between different projects).
Can manage the current tenant’s system roles and role permission configurations.
Can manage sub-account permission allocation of the current tenant system.
Can view the operation logs of the current tenant system. - Early Warning Big Data (Management Cockpit)
-
Support the management of system dictionaries (dynamic management of multi-type early warning dictionaries).
Support system docking app and secret distribution.
Support dynamic control of system menu and permissions.
Support the management of system user departments.
-
測試效果與應用
測試包括Linux系統和Windows系統的c++和Python兩個語言版本的測試
NVIDIA 顯卡環境的測試
華為升騰顯卡/CPU環境的測試
Unbuntu 20.4/22.0/windows 11/10的測試
硬體選購:
- 低配jetson naco(b01替換版本),ubuntu,價格1500元(含外殼),0.5tops,分析實時視頻約4路;一張圖約200ms~300ms耗時;
- 中配jetson orin nano,ubuntu,價格3500元上下,20tops,分析實時視頻約8路;一張圖約100ms耗時;
- 高配jetson orin nano,Ubuntu,價格在5500元上下,70-100tops,分析實時視頻約16路;一張圖約30ms耗時;
- 國產華為晶片orange PI,4G,16核心,價格在1000元左右,分析影片預估在16路;
- 以上價位是基於2024年的市場參考價格,隨時變動,具體查閱官方售價。
- Orange PI Base On Ascend GPU 基於華為升騰GPU的運算測試