ESP32 P4 + Camera + Sensors = ？

[jennazhuxj]

In AI camera development, we realized AI cameras are more than just RGB or depth cameras. "Specialized cameras" like single-point ranging and thermal imaging skip complex image processing yet directly capture key physical data (e.g., temperature, distance), offering a new way to observe the world. This sparked an idea: could we build a lightweight hardware device to make these efficient sensing capabilities more user-friendly, and integrate them directly with camera images or even AI detection results for versatile functions?

The resulting SensorCam P4 is powered by the high-performance ESP32-P4 as its main controller, with functions enabled via pluggable expansion backplanes: Insert a thermal imaging module to display thermal images in real time and mark abnormal areas; insert a laser ranging module to show the precise distance and type of objects; it also supports custom modules—all without the need to reflash firmware.​

We chose the ESP32-P4 because it perfectly meets our needs for data processing and sensor fusion: it features native camera/display interfaces, a 400MHz CPU with AI extension capability, abundant high-bandwidth interfaces, and a mature development environment.​
Currently, we are advancing hardware component selection, device enclosure design, backplane interface development, as well as the development of drivers and UIs for the first batch of modules (camera, laser ranging, thermal imaging). Our goal is to enable the camera to "see temperature" and "see distance".​

We would particularly love to hear your suggestions: Which sensors have the greatest potential when combined with camera images? We will continue to update the project progress on the project page and look forward to interacting with you to optimize SensorCam P4 together!​ https://hackaday.io/project/202943-peek-under-the-hood-how-to-build-an-ai-camera/log/242918-esp32-p4-camera-sensors

[beniroquai]

thanks for sharing! Would that mean the recamera module will move to the esp32-p4 chip?

[jennazhuxj]

@beniroquai reCamera is actually a series of AI cameras in our roadmap. ESP32-P4 might be one of the next-gen AI cameras!
We are testing different potential chips and solutions for the AI cameras and trying to get your suggestions at the same time! Rockchip RV1126B, ESP32 P4 might both be our options!

[beniroquai]

sounds interesting!! Will you also explore different camera modules for the p4? Higher resolution would be great!

[shervinemami]

Will ESP32 P4 allow for much RAM? My guess is that RAM will be the main difficulty for developing AI on ESP32 P4 compared to RV1126B and similar SOCs. For final production-grade solutions, RAM can be low, but for development by customers that are testing one-off projects, having low RAM will cause many difficulties. I don't know the limit for ESP32 P4 RAM but I assume it's very low?

[shervinemami]

Anyway, having a depth image or a thermal image could make computer vision AI much more reliable and possible with low compute, instead of needing a powerful NPU or powerful CPU to perform complex computer vision we can use slow CPU and slow NPU.

[IoTEgr]

The ESP32-P4 itself only carries ~768 KB of on-chip SRAM, which is tiny, but it is designed to address up to 64 MB of external PSRAM. Espressif already ships modules that integrate 16 MB or 32 MB of PSRAM (look for the “R16” / “R32” suffix).
So, for your customers’ one-off development boards you can simply pick a 32 MB module; that is enough RAM to run most lightweight AI PoCs without the pain you’d see on a bare 768 KB MCU. Once the model is proven, you can drop back to a smaller (or no) PSRAM configuration for the final product to save cost.
In short, RAM is not a hard ceiling—just choose the right module for the development phase.

Will ESP32 P4 allow for much RAM? My guess is that RAM will be the main difficulty for developing AI on ESP32 P4 compared to RV1126B and similar SOCs. For final production-grade solutions, RAM can be low, but for development by customers that are testing one-off projects, having low RAM will cause many difficulties. I don't know the limit for ESP32 P4 RAM but I assume it's very low?

[ramkumarkoppu]

Does ESP32-P4 has AI accelerator? Can you please share any more information on this please?

[IoTEgr]

ESP32-P4 lacks a discrete AI accelerator, yet its dual-core 400 MHz RISC-V cores (with FPU + vector extensions), ISP, PPA and hardware codecs deliver ample MIPS for edge CV/ML. Just convert your model to Espressif’s .espdl via ESP-PPQ, then deploy it through ESP-DL—see the https://docs.espressif.com/projects/esp-dl/zh_CN/latest/tutorials/how_to_deploy_yolo11n.html