We apply the deep learning algorithm to the camera in order to implement the real-time object detection and pedestrian tracking in the GPU development board. In addition, it is of the essence to maintain the accuracy rate and the lightweight model can be successfully executed on the PYNQ-Z2 to achieve real-time computing as well.

Use deep learning to detect indoor objects and track predictions. Objects include pedestrians and automobiles. Use compression techniques to train the model to reduce the amount of parameters. We launched the Agile Model. Compared with Tiny-Yolo, the Model Size is reduced by 97.4%, the execution speed is increased by 15FPS, and the embedded platform TX2 reaches 30FPS, and the AP can reach 93.5%.

We provide the state-of-the-art technique of object detection, trajectory prediction, and distance information. It can be applied to anti-collision warning function and increase the convenience of automatic farm equipment, drones and etc. This application can reach our equipment maintenance demands and enhance the quality of public safety issues, such as airports, smart home devices, and security improvement.