Practical usage experiments of high dynamic CMOS sensor IMX174 from Sony in traffic surveillance applications

 

MATRIX VISION is proud to present this sensor as a world’s first in its high-end dual GigE camera family. The camera, called mvBlueCOUGAR-XD104d comes in a black and white and color version, combining this novel sensor with an high speed sensor interface enabling >100 fps at full resolution or 12 bit images with ~65 fps.

 

The camera family is well known for its smart features, such as burst mode buffer, flexible single and dual GigE output modes and a large variety of digital inputs and high current outputs, as well as a video iris and additional universal outputs to drive three motor lenses (focus, zoom and aperture).

 

EMVA1288 overview data shows an very low dark current and noise, combined with an extremely high quantum efficiency and a high end CCD like full well capacity of the 5.8 x 5.8 µm pixel size of >32.000 e^-. This results in a dynamic range of >72 dB and a very high signal to noise ratio of ~45 dB.

 

 

This white paper aims to demonstrate dynamic range and low noise in easy to understand experiments using simply the camera, a laptop, a road, a nice car (J) and ambient light on a sunny day.

Before we start playing let’s have a look at the following screenshot which gives a sample overview of the relevant EMVA1288 data summary:

 

 

 

For this purposely simple demonstration we only used the camera, equipped with a linear polarization filter to get rid of some strong reflections on the car and windshield surface and ambient light. The camera was parameterized in a way, that with 1.5 millisecond exposure time a non-saturated image could be achieved. This can be noticed by the cloud in the sky which is clearly visible. Furthermore we used a 10mm lens and an aperture of f:4. Please also note that due to linear processing and display (no GAMMA processing in the signal path) strong contrasts between the bright and the dark sections in the image are visible.

 

 

The sensor has a built in 12 bit analog front end with extremely high performance, which is key to the overall image quality. Having said that the camera can produce 12 bit images and knowing that a typical computer can only display 8 bit images, MATRIX VISION has introduced in its powerful viewer wxPropView the possibility to bit-shift and clip the image plane by simple commands. This can be compared with a selective digital multiplier of 6dB.

Details are nicely explained here:

 

 

The next image shows the same overview image as before but shifted by one bit to the left and clipped (so the white engine hood stays white).

 

 

A digital zoom into the license plate region is shown in the following image:

The clear and distinct image should not pose any problems on any OCR reading algorithm.

 

 

For traffic surveillance applications, it is of course important to identify not only the car but also the car driver. This can present typically a problem still today because of the lower dynamic range and higher noise of traditional cameras due to the following camera parameterizing dilemma: If the reflective license plate is not to be oversaturated then the driver is almost lost in the dark noise sitting behind the cars IR reflective and darkened windshield. Due to the much higher dynamic and lower noise of the new sensor, we achieve with this new sensor a dramatically better image even without any additional IR flash or filtering.

 

The next screen shots show the driver AOI in the image, using double bit-shift (=12dB local gain and clipping on the left side and triple bit-shift on the right). Of course, this image is resolution-limited but noise is remarkably low. Again these results were achieved with no additional illumination as usually is present in law enforcement applications.

 

 

As an alternative of working and transmitting images in 12 bit, it is possible to compress the dynamic range of the image in the camera using precision LUT-based GAMMA preprocessing.
 

 

The above image shows the car area processed with a standard GAMMA=1/2.2.

This nonlinearity can be compared to a selective amplification: The first 1% input signal is amplified by a factor of 13, 5% signal are still amplified by a factor of 5. The camera is able to process this nonlinearity in real-time (on the fly) with 12 bit accuracy using a 12 bit to 10 bit converting RAM-LUT.

 

It may be also important to see that this new CMOS sensor is free of certain image artifacts like smear (a white vertical line through the image of a CCD at the position of the reflection) or black sun (a bright spot can have a black hole inside), which can be seen on certain other CMOS sensors.

 

The next and last images show a screenshot with quite heavy reflection on the windscreen, but neither of the two artifacts are visible, even not with two times bit shift.

 

 

 

 

A major breakthrough in machine vision applications is achieved by using Sony’s new CMOS global shutter imager IMX174 in MATRIX VISION’s new camera mvBlueCOUGAR-XD104d. Introducing this new camera model, limiting image artifacts are all overcome for challenging applications like but not only restricted to traffic surveillance.

 

Oppenweiler, 22.7.2014

 

Horst A. Mattfeldt