Object recognition on raspberry pi using tensorflow
Loading...
Date
2019-06
Authors
Tan, Hang Yan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Deep Learning (DL) has been widely used in many applications and implemented in commercial and industrial embedded systems. There are a lot of DL applications that will rule the world in the future such as self-driving cars, healthcare, voice search assistants and automatic text generation. TensorFlow is one of the open source DL frameworks. TensorFlow can be used for high performance numerical computations and its flexible architecture allows easy deployment of computation across a variety of platforms. In this project, the accuracy and speed of TensorFlow pretrained models SSDLite MobileNet V2 and SSD Inception V2 implemented in the developed object recognition system on Raspberry Pi 3 Model B+ are evaluated and compared. In this research, COCO 2017 has been used to evaluate the accuracy of SSDLite MobileNet V2 and SSD Inception V2 in terms of 0.7 confidence level on detecting five classes of object. Furthermore, a video input is used to evaluate the speed of SSDLite MobileNet V2 and SSD Inception V2 in terms of Frame per Second (FPS). The accuracy evaluation result shows that the SSDLite MobileNet V2 has a detection score of 69.83% with standard deviation of 3.96% for all five classes. Meanwhile, SSD Inception V2 shows a detection score of 73.72% with a standard deviation of 4.70% for detection of five classes of objects. In addition, the speed evaluation result shows that mean FPS of SSDLite MobileNet V2 is 1.01 with a standard deviation of 0.107. SSD Inception V2 gives a speed evaluation result with mean FPS equal to 0.48 with a standard deviation of 0.041. With these results, it is concluded that SSD Inception V2 performs better than SSDLite MobileNet V2 in the accuracy assessment while SSDLite MobileNet V2 performs better than SSD Inception V2 in the speed assessment.