Towards Real-Time Object Detection on Embedded Systems

Convolutional neural network (CNN) based methods have achieved great success in image classification and object detection tasks. However, unlike the image classification task, object detection is much more computation-intensive and energy-consuming since a large number of possible object proposals need to be evaluated. Consequently, it is difficult for object detection methods to be integrated into embedded systems with limited computing resources and energy supply. In this paper, we propose a pipelined object detection implementation on the embedded platform. We present a comprehensive analysis of state-of-the-art object detection algorithms and select Fast R-CNN as a possible solution. Additional modifications on the Fast R-CNN method are made to fit the specific platform and achieve trade-off between speed and accuracy on embedded systems. Finally, a multi-stage pipelined implementation on the embedded CPU$+$ GPU platform with duplicated module-parallelism is proposed to make full use of the limited computation resources. The proposed system is highly energy-efficient and close to real-time performance. In the first Low-Power Image Recognition Challenge (LPIRC), our system achieved the best result with mAP/Energy of 1.818e-2/(W$\cdot$ h) on the embedded Jetson TK1 CPU$+$ GPU platform.