Rear Overtaking Warning System Implemented by Deep-learning-based Multi-object Detection and Behavior Recognition

• Main Authors: Tseng, Ching-Kan, 曾慶鎧
• Other Authors: Guo, Jiun-In
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/4gdxxg

碩士 === 國立交通大學 === 電子研究所 === 107 === This thesis combines 2D convolution object detection network and 3D convolution behavior recognition network result to achieve rear overtake warning system. On a 2D convolution object detection network, we modify object detection network to let a network with a small amount of parameters generally have a better detection effect. Additionally, we revise the original open source to make this architecture more platform portable than other architectures. On the 3D convolution behavior recognition network, we improve the original architecture to make behavior recognition network with low resolution input have the ability of object localization, which utilize the last layer of convolution layer to learn the overtaking object location in the latest image. Finally, we combine 2D object detection and 3D behavior result together for application purpose. The proposed system is not only developed on PCs but also implemented on the platforms of the embedded systems. After 2D object detection network is modified, it can achieve 480x320 input resolution 5FPS in Renesas RCAR-H3 and 416x416 input resolution 1.5 FPS in Huawei hikey960. Besides, Behavior recognition network can be implemented on Nvidia Tegra-TX2, which reaches 5FPS.