A Fast and High-Performance Object Proposal Method for Vision Sensors: Application to Object Detection
Huawei Liang,Zhiling Wang,Shuhang Tan,Chao Jiang
DOI: https://doi.org/10.1109/JSEN.2022.3155232
IF: 4.3
2022-05-15
IEEE Sensors Journal
Abstract:Use of the object proposal method as a preprocessing step for object detection of vision sensors has improved computational efficiency in recent years. Good object proposal methods require high object detection recall, low computational cost, good localization accuracy, and repeatability. However, existing methods cannot always achieve a good balance of performance. To solve this problem, we propose a fast and high-performance object proposal algorithm. First, we propose a construction method to enhance frequency features that are combined with a linear classifier to learn and generate a set of proposal boxes. Second, we propose a strategy of binarizing frequency features and classifiers to accelerate the calculation. Last, we propose a merging strategy to improve the localization quality of the proposal boxes. Empirically, for the VOC2007 and MSCOCO2017 datasets using the intersection over union (IOU) threshold of 0.5 and 104 proposals, our method achieves 99.3% object detection recall, 81.1% mean average best overlap, and 80% mean repeatability with an average time of 0.0014 seconds per image. The average time is three times faster than the current fastest method, and the mean repeatability is 11% higher than that of the region proposal network (RPN) method. We applied our method to the target detection of autonomous vehicles, and in the experiment with the Oxford RobotCar dataset, we achieved 95.6% detection precision and 91.2% detection recall. This work could provide a new way to improve real-time performance and detection accuracy in the object detection of visual sensors.
Engineering,Computer Science