i.MX95 CMS Demo

This is a suite of libcamera-based camera applications optimized for NXP platforms. This implementation demonstrates a Camera Monitoring System (CMS) on the NXP i.MX95 EVK with dual-camera support, ultra-low latency display, and advanced lens contamination detection.

1. Background: What is a Camera Monitoring System (CMS)?

A Camera Monitoring System (CMS) is an advanced automotive technology that replaces traditional side-view and rear-view mirrors with high-resolution cameras and in-vehicle displays. CMS offers several advantages over conventional mirrors:

• Expanded Field of View: Reduces blind spots and improves scene coverage.
• Improved Aerodynamics: Reduces vehicle drag, improving fuel efficiency.
• Better Visibility: Performs well in low-light conditions and adverse weather.
• Advanced Features: Supports object detection, lane departure warnings, and contamination detection.
• Compact Design: Reduces vehicle width and improves maneuverability.

2. CMS Demo Overview

The CMS demo, built on the NXP i.MX95 EVK, delivers real-time vision performance with dual OS08A20 cameras capturing 1080p at 60 fps. The cameras are frame-synchronized and each drives its own display, achieving an ultra-low latency of just 50–60 ms from capture to output.

Key Features

Feature	i.MX95 EVK
Dual Camera Support	OS08A20
Max Resolution	1080p @ 60fps
Frame Synchronization	Hardware-sync
End-to-End Latency	50-60ms
Dual Display Output	Independent
Preview Modes	EGL, DRM
Video Encoding	H.264/H.265 (VPU)
MP4 Container	Optional
Object Detection	CPU/NPU (SSD)
Lens Contamination Detection	OpenCL-accelerated
Image Processing	G2D Hardware
Zero-Copy DMA Buffer	ISP/GPU/DPU/VPU

Unified DMA buffer shared across ISP/GPU/DPU/VPU — no memcpy, no format conversion. Optimize hardware performance to its fullest potential.

3. Hardware Requirements

Board Name	Quantity	Descriptions
IMX95LPD5EVK-19	1	i.MX95 19x19 EVK evaluation board
IMX-OS08A20	2	OS08A20 camera module
IMX-LVDS-HDMI	1 or 2	If HDMI monitor is connected through LVDS interface
IMX-MIPI-HDMI	1 or 2	If HDMI monitor is connected through MIPI DSI interface

Note: For dual display output, you can use either LVDS or MIPI DSI interfaces, or a combination of both, depending on your display configuration.

4. Software Overview

Information	Value
i.MX BSP	LF-6.18.2-1.0.0

System Architect:

NOTE: FFmpeg is used for MP4 video encoding; it is optional and can be removed.

Optional Dependencies for MP4 Video Encoding

Important: The following FFmpeg libraries may NOT be included in the default Yocto release:

• libavcodec.so.*
• libavformat.so.*
• libavutil.so.*
• libswresample.so.*

Option 1: Deploy FFmpeg Libraries from Yocto SDK (Recommended)

If you have the Yocto SDK installed, copy the libraries to your target board:

Example:

scp ${SDK_PATH}/sysroots/armv8a-poky-linux/usr/lib/libavformat.so* root@ip_address:/usr/lib/

Option 2: Build Without MP4 Support (Raw H.264 Only)

If you only need raw H.264 output and want to avoid FFmpeg dependencies, you can disable MP4 muxing in preview/CMakeLists.txt.

Option 3: Build FFmpeg in Yocto Image

5. Build Instructions

Install Dependencies

source $SDK

Build the Project

mkdir build && cd build

cmake ..

make -j$(nproc)

Deploy and Execute the Binaries

Copy the required binaries to the target i.MX95 EVK:

scp libcamera_CMS          root@ip_addr:/root/
scp preview/libpreview.so  root@ip_addr:/usr/lib/
scp core/libcamera_app.so  root@ip_addr:/usr/lib/
scp ../preview/models/*    root@ip_addr:/root/

Set up the environment and execute the application:

export LIBCAMERA_PIPELINES_MATCH_LIST='nxp/neo,imx8-isi,uvc'
systemctl stop weston*
./libcamera_CMS -n 2 -p 0 -d 0

Reduce the buffer size

The default buffer count is 8. Reduce it to 4 by modifying:

vi /usr/share/libcamera/pipeline/nxp/neo/config.yaml

6. Usage

Basic Single Camera Mode

./libcamera_CMS -n 1 -p 0 -d 0

This runs with one camera using EGL preview, no object detection.

Dual Camera CMS Mode

./libcamera_CMS -n 2 -p 0 -d 0

This runs with two synchronized cameras using EGL preview.

With Object Detection

# CPU-based detection
./libcamera_CMS -n 2 -p 1 -d 1

# NPU-based detection (faster)
./libcamera_CMS -n 2 -p 1 -d 2

Command Line Arguments

Usage: ./libcamera_CMS [options]

Camera Options:
  -n, --num-cameras <N>      Number of cameras (1 or 2, default: 1)
  -p, --preview <TYPE>       Preview type (0=EGL, 1=DRM, default: 0)
  -d, --detector <TYPE>      Detector type:
                               0 = None (default)
                               1 = CPU detection (SSD MobileNet)
                               2 = NPU detection (SSDLite)
                               3 = Dirty detection
Recording Options:
  -r, --record               Enable video recording
  -o, --output <FILE>        Output file (default: output.mp4)
                             For dual camera: output_cam1.mp4, output_cam2.mp4
  -c, --codec <CODEC>        Video codec (h264/h265, default: h264)
  -b, --bitrate <RATE>       Bitrate in bps (default: 8000000)

Other Options:
  -v, --verbose              Enable verbose output
  --help                     Display this help message

7. Lens Contamination Detection (Dirty Detection)

Overview

The dirty detection system uses OpenCL-accelerated image processing to identify lens contamination (dirt, dust, water droplets, etc.) in real-time. It employs a multi-stage filtering pipeline to distinguish actual contamination from normal scene elements like walls, ceilings, or shadows. It divides the image into 64x64 pixel grids, and compute per-grid metrics:

• Contrast: Difference between the brightest and darkest pixels in the grid.

• Gradient: Average edge strength.

• Brightness: Average Y value.

  

Temporal Stability Tracking

The temporal stability filter uses a sophisticated region matching algorithm:

8. Results

Demo Output

The following image shows the CMS demo running in dual-camera mode with DRM preview:

9. Troubleshooting

FFmpeg Library Not Found

Error:

./libcamera_CMS: error while loading shared libraries: libavcodec.so.61: cannot open shared object file

Solution:

1. Copy libraries from Yocto SDK (see "Deploy FFmpeg Libraries" above)

10. Known Issues

Frame synchronization

The sensor driver was modified to enable dual camera frame synchronization. Different BSP versions may require different implementation approaches.

@@ -875,7 +876,44 @@ static int ox05b1s_apply_current_mode(struct ox05b1s *sensor)
 				      sensor->mode->reg_data_count);
 	if (ret)
 		goto out;
+	if (sensor->i2c_client->adapter->nr == 1) {
+       //set the register settings:
+       //Master setting
+		(0x3002, 0x80);(0x3009, 0x06);(0x377e, 0x08);(0x3818, 0x03);
+		(0x3819, 0xc8);(0x381a, 0x04);(0x381b, 0xa0);(0x3823, 0x08);
+		(0x3824, 0x00);(0x3825, 0x20);(0x3826, 0x00);(0x3827, 0x08);
+		(0x3832, 0x22);(0x3834, 0xf4);(0x3842, 0x00);
+	}
+	else {
+       //Slave setting
+		(0x3002, 0x00);(0x3009, 0x02);(0x377e, 0x0a);(0x3818, 0x00);
+		(0x3819, 0x00);(0x381a, 0x00);(0x381b, 0x01);(0x3823, 0x50);
+		(0x3824, 0x03);(0x3825, 0xc8);(0x3826, 0x04);(0x3827, 0xa0);
+		(0x3832, 0x02);(0x3834, 0x04);(0x3842, 0x00);
+	}

Additionally, the J2 pin 2 of both camera modules must be connected together.

11. Release Notes

Version	Description	Date	tag
1.0.0	Initial release	June 2026	cms_v1.0.0

Licensing

CMS is licensed under the BSD_2_Clause.

## Acknowledgments

This project is based on:
- **libcamera-apps** by Raspberry Pi Foundation
- **libcamera** by the libcamera project
- **TensorFlow Lite** by Google
- **OpenCV** by OpenCV Foundation

Special thanks to:
- Raspberry Pi Foundation for the original libcamera-apps
- The libcamera community for the excellent camera framework

## Contact and Support

### Resources

- **NXP i.MX95 Documentation**: https://www.nxp.com/imx95
- **libcamera Documentation**: https://libcamera.org/
- **TensorFlow Lite**:  https://www.tensorflow.org/lite
- **OpenCV Documentation**: https://docs.opencv.org/

---

**Last Updated:** June 2026  
**Version:** 1.0.0  
**Maintainer:** NXP Semiconductor