It can only gather test data and run inference. It is also not powerful enough to do training. The Raspberry Pi 3+ is just barely powerful enough to drive around a track and requires ideal lighting for training and driving. I have been working on upping the game for the Donke圜ar and make AI more accessible to high-school students. It has 128 GPU cores and could be an alternative for the Raspberry Pi for the Donkey Car Performance here was not bad at 3 counts maximum change over the sensor field.Īs with all such tests – further study is needed :-).The Nvidia Jetson Nano is a new single board computer that sells for $100. This isn’t perfect but it does show a general amount of error in readout. Below is a pseudo-colored maximum intensity projection of the 40 frames obtained during the experiment.Ī basic example of read noise on the taps can be seen by looking at the standard deviation over the frames. I’m curious how this may change with a better cooling system. Of the images obtained, a clear pattern can be seen, again with the lens cover on and exposures of 5 seconds. Of additional interest is the noise in the image from the readout taps.

Once the Cooler is disabled at ~ frame 24, the mean continues rising. A drop of ~ 8 counts can be seen after the cooler is enabled. Further tests will include a tightly wrapped enclosure to better thermally sink the complete camera sensor cell and PCB. Not amazing, but the drop in mean image intensity directly matches the enable and disable of the cooler. Images obtained showed corresponding drops in background intensity of ~ 10% for this test. What is quite evident at about the halfway mark, when the cooler is enabled, is that the TEC Cold side goes “dark”, and shortly thereafter the aluminum C mount housing of the camera as well as the exposed PCB begins to cool down. While running the experiment we recorded an IR video to document the reaction.

To test performance, we obtained 40 images with an exposure time of 5 seconds. The camera sits atop a TEC, which sits on an aluminum block, which sits on top of a CPU cooling sink. Next, we used a off the shelf TEC, and a CPU radiator, for the heat exchange. Ideally, we would wrap the entire camera system in aluminum, but for now, this works. This provided a poor but useable thermal coupling to the back of the camera PCB. To start, we simply piled up thermal tape on the backside of the camera. I’ll have more info on the camera itself regarding performance, but my first question was – “can it be effectively cooled? ” Fortunately, having run similar experiments on a Point Grey / FLIR camera, I had most of the components and a basic test method ready to go to figure out the answer. Thanks to some generous industry friends, I finally managed to get my hands on a seemingly unobtanium RPI “High Quality” camera.

Flir camera for raspberry pi