Object Detection under Constrained Hardware Scenarios: A Comparative Study of Reduced Convolutional Network Architectures

Recent advances in deep learning point towards the use of computer vision systems based on Deep Neural Networks (DNNs). However, these network architectures are optimized to be executed in specialized hardware, such as in computers with Graphics Processing Units (GPU). Such hardware is rarely available in embedded computers, for instance, those used by mobile robots, so alternatives must be studied in order to guarantee that mobile systems may still benefit from the applications of deep learning. In this work, we investigate the performance of a vision system for ball detection, based on different configurations of the MobileNet Convolutional Neural Network architecture, under a constrained hardware scenario. By gradually reducing the input size and the number of parameters that compose the neural network and comparing their inference time in an Intel NUC Core i7 mini-PC, embedded in a humanoid soccer robot, we have found acceptable values for the width and resolution multipliers to be used in our soccer ball detection system during a robot-soccer match.