Optimization for object detector using deep residual network on embedded board

Although many studies on deep neural networks for Advanced Driver Assistance Systems(ADAS) have been emerging, using deeper convolutional neural networks (CNNs) implemented on embedded boards or mobile platforms for the road object detection task is challenging. To deploy a deep CNN pipeline, constraints such as computing power, processing time, battery, and memory capacity need to be overcome. To overcome these constraints, this study proposes a smaller residual CNN architecture with 6× fewer parameters and 3× smaller memory that is adopted by the object detector, a single-shot multibox detector. The proposed method was evaluated using the KITTI dataset on the NVIDIA TX1 embedded board and achieves 10.63 frames per second, which is three times faster compared to the well-known model, the VGGNet.