Real time object recognition in videos with a parallel algorithm

Abstract

COmbination of Shifted Filter REsponses - COSFIRE - is a rotation, reflection and scale invariant method for object recognition. Already having exhibited a high degree of accuracy in several applications, its integration into real-time systems has not yet been investigated. This is due to the lack of an implementation which executes at the required frame rate of 30 frames/second. A number of existing methods for object recognition have been accelerated using technologies ranging from hardware implementations to multicore CPU solutions. COSFIRE is an intuitive and flexible neuron-inspired method that is non-application specific. It is also easily trainable due to its ability to derive an invariant filter from a single training image or prototype. It thus merits inclusion in their number. CUDA has been used to accelerate the execution of various algorithms through graphics processing units. Albeit one of several APIs providing access to this massive paralleism, CUDA has been selected due to its abstraction of low-level hardware considerations whilst still exposing sufficent resources for more intensive optimization. Hence, a CUDA implementation of the non-invariant application of a COSFIRE filter to input images has been developed. This implementation displays both data and task parallelism. Data parallelism is made possible through the use of the frequency domain for the execution of the two filtering operations. Task parallelism is actualised using CUDA streams to facilitate concurrent processing of the COSFIRE filter tuples. This venture into task parallelism also expands the horizon for future work investigating the scalability of the implementation, both to more complex COSFIRE filters and the execution of multiple COSFIRE filters. Whilst the target execution time has not yet been achieved, a 9.5x speedup has been observed when compared to the MATLAB implementation of COSFIRE. Thus, this project is the first stepping stone towards a real-time implementation of COSFIRE. The work presented here is in anticipation of further speedup through more aggressive optimisation and more advanced techniques. The three topics central to this project - object recognition, COSFIRE in particular, and Graphics Processing Unit technology - are surveyed, followed by the documentation of the design and implementation of COSFIRE using CUDA. Evalution both in terms of execution times, achieved speedup and the quality of the implementation as indicated by application profiling is also presented. This work is then brought into further context by the discussion of related work and the conclusion’s details of potential future work.