Fully local security behavior analysis and ai device and method

Abstract

To solve the problem that it is difficult to perform real-time analysis in an offline environment or outdoors since a conventional crime prevention system has a communication delay and a privacy leakage risk due to cloud dependence.SOLUTION: In the present invention, a camera video is input, a person or an object is extracted by an arbitrary object detection means, a posture / action feature is estimated by a motion analysis means, and a suspicious action is determined by a machine learning model. The determination result is output as a warning by the speech synthesis, and displayed and recorded on the GUI in real time. To achieve high-speed and safe crime prevention action analysis which is portable by mobile battery drive and does not go through a cloud.

Description

[Technical Field]

[0001] This invention relates to a technology for detecting suspicious individuals and abnormal behavior using real-time behavior analysis with edge AI in the fields of crime prevention and facility monitoring. In particular, it relates to a completely local crime prevention AI device and method that performs object detection and pose estimation from camera images, determines suspicious behavior, and issues audio and visual warnings without using the cloud. [Background technology]

[0002] Conventional security systems rely on cloud analysis, and have issues such as communication delays, privacy risks, and difficulty in use when offline. Additionally, while there are AI-based suspicious behavior analysis systems, there are still few devices that can operate in real time without the cloud, and no integrated systems that combine voice warnings and GUI displays have been disclosed. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Document 1: JP 2019-135678 A (Cloud-based Crime Prevention Analysis System) Relevance: Discloses crime prevention analysis using cloud processing, but does not disclose a fully local configuration. Patent Document 2: JP 2020-056111 A (Behavior Detection AI System) Relevance: Describes behavior detection AI, but does not integrate YOLO + Pose + voice output. [Non-patent literature]

[0004] [Non-Patent Document 1] Non-patent document 1: arXiv:2301.11579 "YOLOv8: Real-Time Object Detection" Related: Discloses high-speed object detection technology, but does not describe behavior analysis or voice output. Non-patent document 2: arXiv:2212.04356 "Pose-based Suspicious Activity Recognition" Related: Relates to pose-based behavior recognition, but does not disclose fully local execution. Summary of the Invention [Problem to be solved by the invention]

[0005] Cloud-based crime prevention analysis has limitations in terms of communication environment and security. The purpose of this invention is to provide a behavioral analysis AI device that operates completely locally, enabling instant detection of suspicious behavior and issuing audio and visual warnings. [Means for solving the problem]

[0006] The present invention comprises the following features: (1) People and objects are captured using a USB camera and detected using the YOLOv8n model. (2) Analyze motion characteristics using a Pose Estimation model (MediaPipe Pose or OpenPose). (3) Suspicious behavior (intrusion, wandering, falls) is detected using a machine learning model. (4) Generate a warning message and output a voice warning in OpenJTalk. (5) Real-time display of warning logs using Tkinter GUI. (6) All processing is performed on a Raspberry Pi 5 + Hailo-8L without going through the cloud. [Effects of the Invention]

[0007] Privacy protection: Cloud-independent to prevent data leakage. High-speed processing: Suspicious behavior detection and voice warning in less than 1 second. Wide range of applications: Can be used in factories, offices, warehouses, schools, unmanned stores, and disaster prevention fields. [Brief explanation of the drawings]

[0008] [Figure 1] Prototype configuration diagram DETAILED DESCRIPTION OF THE INVENTION

[0009] One embodiment of the present invention is a portable crime prevention behavior analysis device that uses a Raspberry Pi 5 and a Hailo-8L AI accelerator, and performs person and object detection, posture analysis, suspicious behavior determination, and audio output on the local device without the need for cloud communication. (1) Hardware configuration Computing device: Raspberry Pi 5, etc. (runs on an OS, has a USB 3.0 port). AI accelerators: Hailo-8L, etc. (converts YOLOv8n model to HEF format for high-speed inference). Imaging method: UVC-compliant USB camera (resolution 1280 x 720 px, 30 fps). Audio Input / Output: USB microphone (for capturing voice commands). Bluetooth or wired speaker (audio warning output via OpenJTalk). Display: 3.5-inch touch display (Tkinter GUI display). Power supply: Can be used outdoors using a 5V / 3A compatible mobile battery. (2) Software configuration Object detection: YOLOv8n model using the ultralytics library. Pose estimation: Use MediaPipe Pose or OpenPose to extract upper body and full body landmarks in real time. Action determination: Extracts movement features from pose information (position of shoulders, waist, and knees). Classify suspicious behavior (intrusion, loitering, falls) using rule-based judgment or random forest / XGBoost models. Speech synthesis: OpenJTalk (Japanese speech library) is used to generate warning voice messages such as "Suspicious behavior has been detected." GUI display: Using the Tkinter library, detection logs are displayed as text, with important warnings highlighted in red. (3) Processing Procedure Imaging: USB camera captures images at 30fps. Object detection: YOLOv8n detects people in video frames in real time. Motion analysis: The detected person area is input into MediaPipe Pose, and posture landmarks (coordinates of shoulders, knees, hips, etc.) are calculated. Suspicious behavior detection: If the difference in height between the shoulders and waist is below a certain level, it is determined to be "crouching / falling." If the movement speed is abnormally slow or if the person remains in a specific area for a certain period of time, it is determined to be "wandering." Warning output: When suspicious behavior is detected, the words "Suspicious behavior detected" are displayed in the GUI log. Generate audio alerts with OpenJTalk and speaker output. Recording: Automatically save log data to CSV or text. (4) Example of operation If wandering behavior is detected for 5 seconds at the entrance or exit of a factory or warehouse, an audio notification will sound saying "Warning: Suspicious person detected" and an alert will be displayed in red text on the GUI. In disaster prevention and nighttime monitoring mode, an announcement can be made saying "Someone has fallen" when a fall is detected. (5) Variations AI model switching: Replace YOLOv8n with YOLOv8s / v8m, or use the optimized version of Hailo. Multi-camera support: Connect two or more USB cameras to expand the monitoring area. In addition to security applications: Can also be used for fall detection and monitoring in nursing homes and welfare facilities. [Example]

[0010] Hardware configuration: Raspberry Pi 5, Hailo-8L, USB camera (1280 x 720px, 30fps), USB microphone, Bluetooth speaker, 3.5-inch touch display. Process flow: Detect people in camera footage using YOLOv8n → Estimate posture using Pose Estimation → Detect suspicious behavior (falling, wandering) using machine learning → Display GUI and output audio alert. Example program: Implemented a script in Python that combines ultralytics.YOLO, mediapipe.pose, and open_jtalk. [Industrial Applicability]

[0011] Security and safety monitoring for factories, warehouses and offices. Detecting suspicious individuals in schools, hospitals, and public facilities. Monitoring people flow and detecting abnormal behavior in unmanned stores and disaster sites. [Explanation of symbols]

[0012] 1. USB camera 2. USB microphone 3. AI accelerator 4. Processing unit Raspberry Pi Reapai 5 5. Bluetooth Speaker

Claims

1. The image captured by the camera is used as input. (1) Extracting a person or object using any object detection means; (2) Estimating the posture or behavioral features of a person using a motion analysis means; (3) Using machine learning models to identify suspicious behavior, (4) outputting a warning by a voice synthesis means based on the determination result; (5) It is characterized by being portable and powered by a mobile battery. A completely local crime prevention behavior analysis AI device.

2. 2. A completely local crime prevention behavior analysis AI device according to claim 1, wherein the determination results are displayed in real time on a GUI display and an alert log is recorded.

3. Using the device according to claim 1 or 2, It is characterized by continuously analyzing camera footage, estimating behavioral characteristics, determining suspicious behavior, and outputting warnings via voice and on screen. A completely local crime prevention behavior analysis AI method.

Images