A Hardware-Accelerated Reduced Dimensionality Multi-Prototype Learning and Recognition System with complementary classifiers

Real-time performance is necessary for many computer vision and machine learning systems in the robotics and automotive domains. Hardware based solutions can meet the demands of real-time applications. Support Vector Machines have demonstrated to be one of the best performing general purpose classifiers, however the approach requires significant simplifications to satisfy hardware constraints. Conversely, associative memory based architectures are inherently hardware friendly and have proven to be a promising approach for high-speed computer vision and machine learning applications, but the accuracies of these systems have not been satisfactory. This paper extends the concept of hardware associative memories and reports an FPGA Accelerated Complementary Reduced Dimensionality Multi-Prototype Learning and Recognition System which is composed of discriminative Reduced Dimensionality Multi-Prototype classifiers. To test the performance, the system is applied to human detection. In addition to outperforming general purpose support vector machine based systems, the complementary approach demonstrated a combined detection time of only 14.9 μs per sample.