Algorytm śledzenia obiektów w obrazie termowizyjnym i jego implementacja w układzie FPGA

W artykule przedstawiono algorytm śledzenia obiektów na obrazach termowizyjnych za pomocą zmodyfikowanej metody SSD oraz propozycję jego implementacji sprzętowej w module programowalnym FPGA. Zastosowanie technologii FPGA pozwoliło na zastosowanie kilku technik przyspieszania obliczeń. Moduły realizujące algorytm zostały zaprojektowane tak, by obliczenia prowadzony były w trybie pipeliningu. Ponadto w celu zwiększenia szybkości działania algorytmu zastosowane zostało zrównoleglenie obliczeń. W artykule opisano architekturę zaprojektowanego systemu przetwarzania obrazów i śledzenia obiektów na obrazie metodą SSD.

In the article the architecture of hardware implementation of SSD tracking algorithm for thermal images is proposed. Object tracking is a process of finding chosen object on the following frame using knowledge about its position in previous frames [1, 3]. Gradient based methods like Sum-of-Squared-Differences (SSD) localize targets by analyzing differences between consequent frames. Finding target movement is performed by searching minimum of cost function in space and time. Cost function in this approach is a sum of squared differences. Sum of squared differences coefficient is a measure of difference between two fragments of images and equals (1). If searched object was detected at point (x, y) in previous frame, finding its location in following frame would mean finding (u, v) for which SSD coefficient is the smallest. The picture fragment centered at (x, y) with size equal to the size of the object is treated as the object model. Point (u, v) will then be a centre of the object that is the most similar to the model. This object in new frame is the one found by the SSD algorithm. SSD object estimation is not always reliable, when object is obscured or noised. To distinct reliable position estimation from noisy one the special SSDVar (2) coefficient was developed. The algorithm to calculate SSD coefficient for set of image fragments was proposed to be implemented in hardware, using parallel computation for every compared image fragments. The architecture of parallelized SSD computation unit is shown on Fig. 4 and Fig. 5. Main parts of computation unit were simulated in Quartus II environment.