Research On Object Detection Algorithm For Indoor And Outdoor Scenes

Nowadays,there is a tendency of computer vision in the field of artificial intelligence.Computer vision is a science that studies ways to make a machine look.This field is extensive and complex.Object detection is part of the tasks of computer vision.On the one hand,object detection is a foundation for more image recognition tasks.On the other hand,object detection has many practical applications.Object detection has also solved some problems after several years of research,but the object detection for natural scenes was difficult to adapt to diverse indoor and outdoor scenesdue to the uncertainty of weather and illumination.Therefore,it is necessary to develop some solutions for object detection in specific indoor and outdoor scenes.In this paper,we explore the performance of the improved object detection algorithm Soft-YOLO applied to the Pascal VOC dataset,the pedestrian and vehicle orientation detection dataset of Longhai City,Zhangzhou City,Fujian Province,and the tableware detection dataset of a factory.In addition,pre-processing solutions are given for the poor detection results in two practical scenes.First,we explore the performance of the algorithm Soft-YOLO,which is obtained by replacing the NMS with Soft-NMS in the YOLO series algorithm.On the Pascal VOC dataset,the IoU and the recall is increased,whichcan explain that the improvement increased the location accuracy on obj ect detection.Subsequently,this paper used the Soft-YOLO algorithm applied to the pedestrian and vehicle orientation detection dataset in Longhai City,Zhangzhou City,Fujian Province.When the NMS in YOLOv3 is replaced by Soft-NMS,the IoU increases from 76.70%to 79.55%.In addition,when Soft-YOLOv3 applied to the tableware dataset,IoU increased from 77.32%to 80.66%.Location accuracy is both improved in indoor and outdoor scenes.In practical applications,the detection of pedestrian and vehicle orientation detection dataset was not accomplished during the rainy day,because the rain brought pixel noise to the images.When the image is preprocessed by the de-rain algorithm DID-MDN,the image quality and the detection performance are improved.In different scenes,the results of tableware detection are less robust due to different illuminationsituations.We utilized Gamma correction for preprocessing in the end.In the dark dataset,IoU increased by 3.14%after Gamma correction;in the light dataset,IoU increased from 77.71%to 80.30%after Gamma correction,and both increased mAP,indicating that the tableware detection performance was improvedafter the Gamma correction.