Pedestrian Detection Under Occlusion Based On YOLOv3 Algorithm

With the development of artificial intelligence and computer hardware,pedestrian detection technology has been applied in autonomous driving,intelligent video monitoring,intelligent robots and other fields.Among them,target detection algorithm YOLOv3 based on convolutional neural network has the advantages of better real-time detection and higher detection accuracy,which is widely used in pedestrian detection field.However,there is still a problem that pedestrians block each other in dense scenes,which makes it difficult for YOLOv3 to accurately detect the blocked pedestrians,and it is also difficult for YOLOv3 to accurately extract the whole body features of pedestrians due to the occlusion of pedestrians.Aiming at the problem that it is difficult for YOLOv3 to accurately detect blocked pedestrians due to mutual occlusion between pedestrians,this paper proposes a YOLOv3 pedestrian detection algorithm based on border rejection loss function and improved non-maximum Suppression(NMS).(1)In the training process,YOLOv3 used the border positioning loss function to calculate the positioning error between the pedestrian prior prediction box and the marker box,and corrected the error in the regression process,so that YOLOv3 could accurately locate the pedestrian.However,when the pedestrians block each other,the pedestrian prior prediction box and the marker box overlap each other,resulting in the algorithm to judge the two pedestrians blocking each other as single pedestrians,affecting the border positioning loss function,so that the network can not accurately identify the pedestrians blocking each other.In this paper,a set of loss functions is added to represent the error between the target pedestrian and the neighboring pedestrian.In the training process,the influence of the neighboring pedestrian on the target pedestrian is reduced,so that the network can accurately identify the pedestrians blocking each other.(2)In pedestrian detection under dense scenes,when pedestrians block each other,the pedestrian prior prediction box with a low confidence score will be judged as a false pedestrian prediction box by non-maximum suppression,resulting in the phenomenon of missing detection.Based on this,this paper uses linear function to improve the confidence score function in the non-maximum suppression,weakens the suppression of the non-maximum suppression on the prior prediction box with poor detection effect,and thus reduces the phenomenon of missing detection.In addition to shielding between pedestrians,a large number of pedestrians are blocked by objects in pedestrian detection,which makes it difficult for YOLOv3 to extract the whole body features of pedestrians and causes pedestrians to be missed.Based on this,this paper adds a prediction network branch after the trunk network of YOLOv3 to detect the head andshoulder position that is difficult to be blocked in the actual scene,and USES the head and shoulder prediction results to enhance the robustness of the overall prediction result of pedestrians,and improve the detection accuracy of pedestrians when they are blocked by objects.Caused by a lack of YOLOv3 multi-scale detection ability at the same time increased the pedestrian is blocked by the object scenario testing difficulty,therefore this article through deepening YOLOv3 backbone network,increasing the output characteristics of level,increase the diversity of the anchor box dimension method to strengthen the network under the different distance obscured the pedestrian detection ability,improve the multi-scale detection capability of the algorithm.Finally,the proposed algorithm is verified by pedestrian test set and compared with the original YOLOv3 algorithm.Experimental results show that the proposed algorithm can improve the pedestrian detection performance in the occlusion scene.