居家 异常 行为 监测 系统 开发

@lakshithap1

With my experience in training models with deep learning and traditional computer vision algorithms and my well-documented proficiency in data cleaning, feature engineering, you can expect clean and readable codes along with APIs that facilitate smooth model inferences and efficient data insights. Regarding your demand for a functional prototype to promptly demonstrate its value, I am proficient in creating dynamic and user-friendly applications using various frameworks like React Native, Flutter etc., ensuring that deploying your system locally or on the edge is hassle-free. Emphasizing on efficient solutions, I'm also experienced with dockerized Linux environments that expedite iterative developments while maintaining the system's robustness. Lastly, beyond just ticking off items from a job description list, I feel committed to the overall success of your project. Working on projects related to healthcare has provided me insight into sensitive demands of geriatric care, rehabilitation and assisted living systems which perfectly overlap with your requirements. Having also worked on collaborative projects emphasizes my ability to seamlessly communicate complex technical details. Let's join hands to swiftly transform this safety solution for domestic environments into reality.

@softdev0924

Dear Client~ I read your description carefully, and I am very interested in your project. I have rich experience with computer vision, and I know what you really want. Also, I am fluent in device development, such as Raspberry Pi, Jetson Nano, Jetson TX1, TX2, etc. I developed a China License Plate recognition engine and deployed it on a Raspberry Pi and Jetson TX2. I look forward to connecting with you. Best Regards. Roovee.

@mengw3

您好， 本项目本质上是一套「多模态感知 + 实时行为建模 + 健康风险推理」的边缘智能系统。技术难点集中在三处：跌倒/静止检测的高召回率与低误报率的平衡；个体行为节律的在线学习与偏离预警；毫米波微动信号的呼吸/心率特征提取与个人基线管理。这三点我均有实际项目经验。 技术方案概要 【传感器选型】 主力感知采用广角 RGB 摄像头（≥1080P，视角≥120°）+ 24GHz/77GHz 毫米波雷达组合：摄像头负责姿态与行为识别，雷达负责穿墙呼吸/心率提取与静止状态判断，互为补充，同时规避纯视觉方案在弱光/遮挡场景的盲区。门磁 + PIR 作为低功耗辅助传感层，捕捉出入节律与如厕频率。 【算法栈】 - 跌倒检测：YOLOv8-Pose 实时骨骼关键点提取 → 双流 LSTM（骨骼角度序列 + 质心加速度）进行时序分类，公开数据集（UR Fall / OOPS）预训练 + 小样本迁移微调，目标精度 ≥95%，误报率 ≤2%。 - 长时静止/失能：基于前景差分 + IOU 追踪的无动作时长计数器，阈值可配置（默认 15 分钟触发一级告警）。 - 行为节律建模：以天为粒度提取出入时间、如厕次数、家电操作序列，构建个人基线分布（高斯混合模型），Isolation Forest 检测离群事件，支持 7 天冷启动后上线。 - 生理体征：77GHz 雷达 I/Q 原始数据 → 带通滤波（呼吸 0.1–0.5Hz，心跳 0.8–2Hz）→ FFT 峰值提取，与历史基线比较后输出风险等级。 【系统架构】 边缘端（Jetson Orin / x86 NUC）运行推理服务，Docker 容器化部署，FastAPI 提供本地 REST 接口；告警事件通过 MQTT 实时推送至云端后台；前端告警看板采用 React + WebSocket 实现秒级刷新。 交付清单与周期 | 阶段 | 内容 | 周期 | |------|------|------| | P1 | 传感器选型报告 + 系统架构设计文档 | 第 1–2 周 | | P2 | 跌倒检测 & 静止失能模块 + 推理 API | 第 3–5 周 | | P3 | 行为节律建模 + 生理体征提取模块 | 第 6–8 周 | | P4 | 功能性原型（本地运行 + MQTT 推送 + 告警看板） | 第 9–10 周 | | P5 | 技术文档、测试用例、KPI 报告、部署验收 | 第 11–12 周 | 总周期约 12 周，每两周提交阶段性成果并同步进展，支持需求微调。 为什么选择我 我具备 OpenCV / PyTorch / TensorFlow 全栈开发能力，熟悉 mmWave SDK（TI IWR 系列），在人体姿态估计与时序异常检测方向有完整落地经验，熟练使用 Linux + Docker 环境进行快速迭代。曾参与养老机构跌倒监测与智慧社区异常行为预警项目，理解真实场景中的数据噪声、伦理隐私与稳定性挑战，能交付生产级代码而非实验室原型。我的主页有通过摄像头识别垃圾并进行分拣的视频，欢迎查看。