Parallel spatio-temporal convolution scheme oriented for hardware real-time implementation

A new parallel approach to the 3D convolution of a series of grey-scale images is presented. An original hardware-oriented 2D convolution scheme is elaborated and extended into the temporal domain. The complexity of the scheme is determined and analysed with respect to the number of non-zero coefficients in a convolution kernel. The parallelisation of convolution operations is analysed and some results are presented. Serial and parallel variants of 2D convolution engines are proposed and their time-cost trade-offs are given. The flexibility of the solution with regard to the scalable size of the kernel is highlighted.