Toward real-time object detection on heterogeneous embedded systems

Abstract

low power consumption and high efficiency of heterogeneous systems improves processing power and enables the implementation of real-time applications. Deep learning, as one of the hottest topics of today, plays an important role in solving difficult problems such as machine vision. The use of traditional methods for solving visual machine problems requires the engineering of features by humans, which makes it difficult to create a comprehensive model for a problem. The use of revolutionary deep learning in the machine vision, which along with the embedded systems can be useful in many today's issues. Convolutional neural networks have shown a high degree of efficiency in the task of categorizing images and detecting objects. An important feature in neural networks is the intrinsic parallelism of its structure, which results in the use of embedded heterogeneous systems that can provide excellent performance in the implementation of neural networks. Implementing real-time objects detection systems in enclosed environments with limited computing resources and memory is challenging. This paper presents a method for implementing the MobileNet-SSD object detection system on the Jetson TK1, which attempts to improve performance by changing the network's convoys and dividing tasks between the central and the graphics processor.