An efficient hardware implementation of CNN-based object trackers for real-time applications

The object tracking field continues to evolve as an important application of computer vision. Real-time performance is typically required in most applications of object tracking. The recent introduction of Convolutional Neural network (CNN) techniques to the object tracking field enabled the attainment of significant performance gains. However, the heavy computational load required for CNNs conflicts with the real-time requirements required for object tracking. In this paper, we address these computational limitations on the algorithm-side and the circuit-side. On the algorithm side, we adopt interpolation schemes which can significantly reduce the processing time and the memory storage requirements. We also evaluate the approximation of the hardware-expensive computations to attain an efficient hardware design. Moreover, we modify the online-training scheme in order to achieve a constant processing time across all video frames. On the circuit side, we developed a hardware accelerator of the online training stage. We avoid transposed reading from the external memory to speed-up the data movement with no performance degradation. Our proposed hardware accelerator achieves 44 frames-per-second in training the fully connected layers.