Abstract: As traditional object detection methods are generally applied to object detection systems with low accuracy, a novel object localization methodology is proposed to boost localization accuracy of state-of-the-art object detection systems. Firstly, the initial set of candidate boxes is confirmed. Secondly, conditional probabilities are assigned for each row and column of the search region, in which these probabilities provide useful information regarding boundaries?location?of?object inside searching region. According to the probabilities, the paper builds in-out probabilities, border probabilities and combined probabilities three models to return the bounding box of an object of interest inside this region to achieve object localization. Finally, combined with the localization model, a convolutional neural network architecture is adopted to train and detect objects in combination with localization model. The experiments for different IoU threshold on PASCAL VOC and COCO datasets are carried out. Experimental result show that the proposed method exhibits better detection accuracy in comparison with traditional methods, and could be applied to the state-of-the-art object detection systems. Furthermore, sliding windows method is applied to generate initial set of candidate boxes, not only simplifies the process of confirming the initial set, but also maintains high detection accuracy, which illustrates the proposed method is completely independent of the traditional bounding box regression paradigm.

Key words: object detection, convolutional neural network, object localization, conditional probabilities, bounding box

摘要： 针对传统目标探测方法多应用于低定位精度系统的情况，提出一种目标定位探测方法，以增强目标探测系统的定位精确度。确定候选框初始集；计算给定搜索区域的每行每列元素的条件概率，这些概率提供目标边界框位置的有用信息，根据概率情况，分别建立内外模型、边界模型和混合模型以返回感兴趣目标的边界框，实现定位；结合定位模型，利用卷积神经网络对目标进行训练探测。通过对PASCAL VOC和COCO数据集不同IoU阈值情况的实验，结果表明，与传统的方法相比，提出方法具有更高的探测准确率，可应用于高级目标探测系统。同时，利用滑动窗的方法确定候选框初始集，说明提出方法完全独立于传统的边界框回归方法。既简化了初始集的确定过程，同时保持较高的探测准确率。

关键词: 目标探测, 卷积神经网络, 目标定位, 条件概率, 边界框