Faster CNN-based Object Detection with Adaptive Network Selection on Embedded System

Due to the dramatic growth of the amount of video data on the Internet, a need arises for processing the data at cameras and reducing network traffic. Cameras may be deployed in many energy-constrained systems, such as drones, mobile phones, surveillance cameras using renewable energy. These systems have limited computational capability for running state-of-the-art computer vision applications using the convolutional neural network (CNN) for object detection. Such methods divide the input images into equal-size cells and predict bounding boxes by regression on each cell. They also provide the trade-off between accuracy and speed by varying the input image sizes. When smaller images are used, object detection is faster but the methods may fail to detect some objects. This thesis investigates the relationships between the input image sizes to the neural networks and the accuracy of the detection and proposes an adaptive method that can dynamically adjust the input sizes to CNN for faster object detection with negligible degradation of accuracy. Experiments demonstrate up to 23.8% of runtime reduction with 0.2% average recall improvement and no average precision loss compared with YOLO v2 with fix input size