Spatiotemporal feature extraction algorithms are widely used in many image processing and computer vision applications.
They are favored because of their robust generated features. However, they have high computational complexity. Parallelizing
these algorithms, in order to speed their execution up, is of great importance. In this paper, we propose new parallel
implementations, using GPU computing, for the two most widely used spatiotemporal feature extraction algorithms: scaleinvariant
feature transform and speeded up robust features. In our implementations, we solve problems with previous
parallel implementations, such as load imbalance, thread synchronization, and the use of atomic operations. Our implementations
speed up the execution by simultaneously processing all the work of each stage of the two algorithms, without
dividing that stage into smaller sequential ones. The allocation of the threads in our implementations further allows them to
increase the occupancy of the GPU streaming multiprocessors (SMs). We compare our presented implementations to
previous CPU and GPU parallel implementations of the two algorithms. Results show that the proposed implementations
could do all the processing in real time with high accuracy. They further achieve higher speedup, frame rate, and SM
occupancy than the previous best-known parallel implementations of the two algorithms. |
