Real-time Pedestrian Classification System Using Deep Learning on a Raspberry Pi Cluster

Abstract

Convolutional neural networks (CNN) are commonly used for object classification. However, CNN is computationally expensive and can have performance issues in real-time applications. The objective of this research is to overcome these disadvantages through efficient design, implementation and deployment of CNN on a Raspberry Pi (Rpi) cluster for real-time pedestrian classification. Through the feasibility test, by running CNN classification on one Rpi 3 Model B, the processing speed of approximately 1 FPS was obtained. This is far from the human reaction time requirement which is set to be less than 0.5 sec. In this thesis, two solutions are proposed. First, architectural design, implementation and experimentation with a cluster composed of 3 RPis to meet the two main requirements. Second, tweaking and optimizing the design of the CNN itself. Through the combination of the two solutions, we could achieve the near real-time classification performances which are 0.16 seconds per image, 79.46% accuracy and false negative rate of the classification results is only 4.08%.