We implemented Z.Zivkovic improved Gaussian Mixture Model (GMM) on a Jetson Tk1 Embedded GPU (running on linux, Ubuntu 14.04), for the task of detecting flying birds in sequences of sky images.
The algorithm achieves real-time performance with an average speed of 8.8 fps (~480fps with a GTX 950 dedicated GPU = dedicated GPU:s can have much better performance)
The video clip we have uploaded here shows the segmented frames of 4 different input videos recorded at 8 fps with a resolution of 1920x1374. We have selected the GMM- parameter values such that they provide high segmentation accuracy for many various scenes, including the ones in those 4 different input videos (Video 1 - 4). Moreover, the algorithm manages to suppress false positive pixels in frames that do not include a flying object. Thus, it can reliably detect flying objects.
PLEASE VISIT OUR PAPER (link is given at bottom) FOR MORE DETAIL AND TO SEE THE SEGMENTATION RESULTS IN TERM OF PRECISION, RECALL, SPECIFICITY AND ACCURACY
We describe thoroughly the challenges associated with this work, how the GMM- parameter values affects the classification of the background and foreground pixels and how it can be improved.
WE DESCRIBE STEP BY STEP ON HOW TO COMPILE AND RUN OPENCV AND CUDA APPLICATIONS IN LINUX ENVIRONMENT (UBUNTU) USING CMAKE. WE ALSO DESCRIBE THE NECESSARY STEPS IN ORDER TO IMPLEMENT OUR IMPLEMENTATION ON GPU (ONLY FOR THE JETSON TK1).
Many thanks to V.Pham et al. for the original GMM implementation: www.codeproject.com/Articles/142859/Extended-GMM-for-Background-Subtraction-on-GPU
Our revised implementation code files, including video files, ground truth images can be accessed here: www.dropbox.com/sh/yv3q6tjk49x3h3x/AACqguSzf5AhIseAKZH1dVOpa?dl=0