Geometric Constrained Real-time Visual Object Detection

Object detection is the basis of many computer vision tasks and have been widely used in autonomous driving,video surveillance and image retrieval.In this paper,the one-stage real-time object detection methods and the geometric constrained methods are further studied.The main contributions are as follows.Problems of one-stage real-time object detector in detecting small objects is investigated and solved without sacrifice the speed.Two problems are observed for one-stage object detector in detecting small objects:(1)Anchor box mechanism only obtains the weak translation invariance for small object detection.(2)Classification and Localization modules are separated.For the first problem,if a small object moves slightly in the image,the detection result will change greatly.Circularly shift the feature maps for small objects prediction is proposed to solve the weak translation invariance problem.Meanwhile weak coupling of classification and localization modules causes the most precisely localized detections may not have the highest confidences,and IoU prediction module and Smooth Non-maximum Suppression are utilized to solve it.IoU-Prediction predicts localization accuracy for detection and is combined with classification information as new confidence of detection.Smooth Non-maximum Suppression averaging the accurately localized detections with low confidences and detection with highest confidence to get more accurate localization result.Lastly,the proposed small-objectness sensitive one-stage object detector is expanded to detection in videos by using trajectory hypothesis to predict the prior locations of objects and selectively reduce the confidence threshold.Deep learning based object detection methods suffers from poor interpretability.Geometric Constrained Loss and Suppression are proposed to introduce geometric constraints into deep network in this thesis.In scenarios where the camera pose does not change greatly like autonomous driving,the camera geometric relationship remains unchanged for a giving class of object.Take pedestrian detection in autonomous driving as an example,the relationship between the height of the pedestrian and its location in image plane is deduced based on the law of perspective projection.Geometric Constrained Loss decreases the contributions of samples that violet the geometric constrained theorem to the classification loss.Meanwhile,Geometric Constrained Suppression lower confidence scores of detections that violate the geometric constrained theorem.By utilizing geometric constrained method on both one-stage and two-stage object detectors,the proposed methods achieves great performance on two object detection datasets.