Implementation of Deep Learning-based Hierarchical Object Detection System for High-Resolution Images
Makoto Sugaya, Yusei Horikawa, Kazuma Mashiko, Tomoya Minagawa, Tetsuya Matsumura · 2022 IEEE 11th Global Conference on Consumer Electronics (GCCE) · 2022
We propose a new deep learning-based hierarchical object detection algorithm for high-definition vision sensors for the goal of controlling autonomous vehicles and drones. The proposed algorithm is implemented based on the YOLOv5 network. Wide area detection using a reduced image and local area detection using the original image are performed hierarchically for a Full-HD image. When combining wide and local area detection, by applying the confidence score extracted by wide area detection to the original image selection, both the detection accuracy and detection numbers are improved compared with the conventional model. In evaluations with the 2K video data set VisDrone, our algorithm achieves two times the number of detections and an improved mean average precision (mAP) of about 3% compared to the conventional method. We also implemented this system on a small, low-cost Jetson Nano board and confirmed that real-time operation is possible at approximately 3 frames per second. Since this hierarchical object detection algorithm supports resolution scaling, it can also be applied to super-high-definition videos such as 4K/8K.