Search

Results of Video Stitching

m202camera360
Jun 10, 2016
2 min read

Similar to Image stitching, Once the image transformation mask is calculated and used for stitching video frames as well. A seam function is also applied to diminish the effect of image boundaries and exposure differences. However, exposure differences can't be removed 100% unless the same exposure value was programmed at the time of capturing the images.

However, This method performs extremely well with offline video stitching, as there is sufficient time for processing the stitched video. We tried to explore the real time capabilities of the implementation. To evaluate a real time capabilities we need to understand that transformation of video frames, stitching and seam function is going to be a burdensome time intensive process when there are 6 cameras, each giving a throughput 720p HD videos. Hence the resultant fps drops near a 2fps on a Intel® Core™ i7-6700HQ Processor. An attempt to do the same was tried using GPU. But in order to transmit Matrix transformation mask to CPU, it needs I/O calls which again slows down the overall system throughput.

However, if we reduce the computation by removing one step at a time, we can see the fps can be increased. For the following video, we are refrained from applying seam function. The fps achieved immediately increased to around 5fps.

And if we refrain from doing Matrix transformation as well, then even Below image shows stitching results (~10 fps), which is as good as the individual throughput of individual camera(12 fps).

The same algorithms were implemented on NVIDIA's JETSON TX1 embedded platform and we got 0.5fps for real time stitching with rotation and seam function.