indoor positioning

📍Focused on Spatial Intelligence and AI Agent research. We build custom localization systems, apps, and algorithms. DM for business or tech inquiries.

[indoor positioning @IndoorLocalize]

@GoogleCloud_ME 啥时候发布，

[indoor positioning @IndoorLocalize]

https://t.co/O45Bd6pJfS Teach a Camera in One Sentence: A Local AI That Spots Falls, No Code, No Training。 Lingmou (灵眸) is a real-time camera monitoring system I built on a local vision decision model. Connect an RTSP stream, and the model watches continuously. Each judgement takes 3 frames from the last 2 seconds, so it can answer motion questions like “is the person moving?”, not just “what’s in this picture?” You write rules as plain-language questions: yes/no, pick one, or graded score. The model returns a probability, not just a verdict. Test a rule on the current frame before saving, and it goes live the moment you save. No code, no model training. To keep false alarms down, every rule has a threshold plus a debounce (N hits in a row before it fires), and a severity level from show-only to a red alert with sound. All rules are answered in a single request, and everything runs on my own PC, so the video never leaves the machine. Faces are blurred in this demo.

[indoor positioning @IndoorLocalize]

Teach a Camera in One Sentence: A Local AI That Spots Falls, No Code, No Training。

Lingmou (灵眸) is a real-time camera monitoring system I built on a local vision decision model. Connect an RTSP stream, and the model watches continuously. Each judgement takes 3 frames from the last 2 seconds, so it can answer motion questions like “is the person moving?”, not just “what’s in this picture?”

You write rules as plain-language questions: yes/no, pick one, or graded score. The model returns a probability, not just a verdict. Test a rule on the current frame before saving, and it goes live the moment you save. No code, no model training.

To keep false alarms down, every rule has a threshold plus a debounce (N hits in a row before it fires), and a severity level from show-only to a red alert with sound. All rules are answered in a single request, and everything runs on my own PC, so the video never leaves the machine. Faces are blurred in this demo.

Use cases: elder care, child care, store watch. What would you want your camera to watch for? Tell me in the comments.

#ComputerVision #AIAgents #AIcamera #LocalLLM #edgeAI #Decisionmodel

[indoor positioning @IndoorLocalize]

@tokufxug https://t.co/CQHtyFP53a I also built one using a decision model, without the need for coding or model training.

[indoor positioning @IndoorLocalize]

Teach a Camera in One Sentence: A Local AI That Spots Falls, No Code, No Training。

Lingmou (灵眸) is a real-time camera monitoring system I built on a local vision decision model. Connect an RTSP stream, and the model watches continuously. Each judgement takes 3 frames from the last 2 seconds, so it can answer motion questions like “is the person moving?”, not just “what’s in this picture?”

You write rules as plain-language questions: yes/no, pick one, or graded score. The model returns a probability, not just a verdict. Test a rule on the current frame before saving, and it goes live the moment you save. No code, no model training.

To keep false alarms down, every rule has a threshold plus a debounce (N hits in a row before it fires), and a severity level from show-only to a red alert with sound. All rules are answered in a single request, and everything runs on my own PC, so the video never leaves the machine. Faces are blurred in this demo.

Use cases: elder care, child care, store watch. What would you want your camera to watch for? Tell me in the comments.

#ComputerVision #AIAgents #AIcamera #LocalLLM #edgeAI #Decisionmodel

[indoor positioning @IndoorLocalize]

@GitHub_Daily https://t.co/CQHtyFP53a 我也复现了一个

[indoor positioning @IndoorLocalize]

Teach a Camera in One Sentence: A Local AI That Spots Falls, No Code, No Training。

Lingmou (灵眸) is a real-time camera monitoring system I built on a local vision decision model. Connect an RTSP stream, and the model watches continuously. Each judgement takes 3 frames from the last 2 seconds, so it can answer motion questions like “is the person moving?”, not just “what’s in this picture?”

You write rules as plain-language questions: yes/no, pick one, or graded score. The model returns a probability, not just a verdict. Test a rule on the current frame before saving, and it goes live the moment you save. No code, no model training.

To keep false alarms down, every rule has a threshold plus a debounce (N hits in a row before it fires), and a severity level from show-only to a red alert with sound. All rules are answered in a single request, and everything runs on my own PC, so the video never leaves the machine. Faces are blurred in this demo.

Use cases: elder care, child care, store watch. What would you want your camera to watch for? Tell me in the comments.

#ComputerVision #AIAgents #AIcamera #LocalLLM #edgeAI #Decisionmodel

[indoor positioning @IndoorLocalize]

@xiaomovps 吃灰一年了，🤣，接到扣子，聊天对话。孩子玩了几天就丢到一旁了

[indoor positioning @IndoorLocalize]

Teach a Camera in One Sentence: A Local AI That Spots Falls, No Code, No Training。

Lingmou (灵眸) is a real-time camera monitoring system I built on a local vision decision model. Connect an RTSP stream, and the model watches continuously. Each judgement takes 3 frames from the last 2 seconds, so it can answer motion questions like “is the person moving?”, not just “what’s in this picture?”

You write rules as plain-language questions: yes/no, pick one, or graded score. The model returns a probability, not just a verdict. Test a rule on the current frame before saving, and it goes live the moment you save. No code, no model training.

To keep false alarms down, every rule has a threshold plus a debounce (N hits in a row before it fires), and a severity level from show-only to a red alert with sound. All rules are answered in a single request, and everything runs on my own PC, so the video never leaves the machine. Faces are blurred in this demo.

Use cases: elder care, child care, store watch. What would you want your camera to watch for? Tell me in the comments.

#ComputerVision #AIAgents #AIcamera #LocalLLM #edgeAI #Decisionmodel

[indoor positioning @IndoorLocalize]

@itwanger 现在是云边协同，

[indoor positioning @IndoorLocalize]

https://t.co/VX9zcJAVGD

[indoor positioning @IndoorLocalize]

@tokufxug 在中国已经大量普及了

[indoor positioning @IndoorLocalize]

@humanexelab 多元融合，imu，地图约束等

[indoor positioning @IndoorLocalize]

项目已开源，GitHub 地址放在文末。

最近做了一个室内定位的小实验：

做过 BLE / WiFi 指纹定位的人应该都知道，真正麻烦的往往不是定位算法，而是前期采数据。

一个停车场、商场、办公楼，如果每个位置都人工采 RSSI，建库和后期更新成本都很高。

所以我做了一个尝试：

只采集部分区域的 BLE 指纹，然后用扩散模型自动补全剩余区域。

整体思路很像图像 Inpainting：

少量真实 RSS 数据
→ 随机 Mask 模拟缺失区域
→ 数据增强
→ Conditional Diffusion
→ 自动生成完整 BLE 指纹地图

项目里用了：
✅ 只采集部分区域，降低人工采样量
✅ 面向少样本 / 稀疏采样场景
✅ 扩散模型自动补全缺失 RSS 指纹
✅ 数据增强缓解训练样本不足
✅ 多种随机 Mask 模拟不同缺失区域
✅ DDPM 训练 + DDIM 快速推理
✅ RePaint 保证已采集区域数据不被修改
✅ U-Net + Attention
✅ 输出 RMSE / MAE / CDF 等完整评估结果
✅ PyTorch + HuggingFace Diffusers 实现

目前实验结果：
未知区域 RMSE：5.33 dBm
MAE：3.86 dBm
50% 补全点误差 < 2.9 dBm

我觉得这个方向真正有意思的地方，不只是“扩散模型也能做信号地图”。

更重要的是：

如果以后只需要采 20%、30% 的点，就能恢复出可用的完整指纹地图，室内定位的建库成本会下降很多。

接下来我准备继续测试：
不同采样比例
→ 指纹地图补全误差
→ 最终定位误差
看看“少采多少数据，定位精度还能保持多少”。

项目代码已经整理，适合：

室内定位研究
BLE / WiFi 指纹定位
少样本 / 稀疏采样研究
扩散模型实践
毕业设计 / 科研实验
定位算法二次开发

如果你也在做停车场、商场、工厂、医院等室内定位场景，可以一起交流。
开源地址：https://t.co/DOXjZoIk9j

#室内定位 #BLE定位 #WiFi定位 #扩散模型 #DiffusionModel #PyTorch #人工智能 #深度学习 #科研 #开源项目 #毕业设计 #算法工程师

[indoor positioning @IndoorLocalize]

@humanexelab 生成rssi 指纹，根据knn 或者深度学习匹配位置，是有设备差异的，大佬是懂的人，

[indoor positioning @IndoorLocalize]

@humanexelab 扩散模型生成，误差几db ，对，有漂移。

[indoor positioning @IndoorLocalize]

熬夜vibecoding 了一个demo,文搜视频，图搜视频，直接定位到目标画面，不用再拖视频进度条人工看画面了。 (比如车子被偷去物业监控室，一帧帧看画面，很痛苦[囧])用于智慧安防，巡检场景。 还能接入大模型做多模态rag #Multimodal #MultimodalRetrieval

[Sundar Pichai @sundarpichai]

Introducing EmbeddingGemma 2, a new open multimodal model that sets the standard for on-device efficiency.

- our first open, natively multimodal embedding model
- handles text, code, image, video, and audio tasks within a lightweight, modular 740M parameter form factor
- ideal for offline, privacy-first RAG when paired with Gemma 4
- outperforms some specialist models more than twice its size

Weights available now on Hugging Face.

[indoor positioning @IndoorLocalize]

室内定位开源项目 #wifi定位 #wificsi #indoorposition #indoorlocalization https://t.co/NQtu0t1QZQ

[indoor positioning @IndoorLocalize]

手机视觉定位，多元融合，不需要部署任何设施， 实现停车场找车位。终于跑通了最小mvp. Localization for Indoor Parking Spot Finding #找车位 #视觉定位 #室内定位 #VisualLocalization #IndoorPositioning #AutoParking

[indoor positioning @IndoorLocalize]

Mobile Visual Localization, Multi-sensor Fusion, Infrastructure-free
Indoor Mobile Visual Localization #MobileVisualLocalization #MultiSensorFusion #InfrastructureFree #IndoorLocalization #ComputerVision #SLAM

[indoor positioning @IndoorLocalize]

现在搞创新太容易了，之前有个电厂朋友问如果做ai agent 设备调度，我调研了一下，然后再整合这个方案，能发一篇很好的论文了。具体方案 https://t.co/5JqjqDy1wX

[Satya Nadella @satyanadella]

Introducing Microsoft-Decision-1, our new model for fast decision-making.

It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality.

We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.

[indoor positioning @IndoorLocalize]

@0x0SojalSec 很早的模型了