Research On Fast Object Detection Algorithm In Complex Environment

Object detection is one of the important research directions in the field of computer vision,and its aim is to locate and classify objects in images.Object detection has been widely used in many fields,for example,intelligent monitoring,face recognition,and unmanned driving.Especially,in the field of intelligent security scenarios,the social public security and maintain social stability will be guaranteed if it is used.All of these,inevitably,have a profound impact on people life.Therefore,it is very necessary to continue researches about the object detection.Although there have been many excellent object detection algorithms,the disadvantage of slow detecting still exists in some algorithms,causing not keeping up with the playback speed of the video stream and missing important frames detection.In addition,with a complex environment,it became difficult to detect the object since the uneven lighting,fuzzy object,occlusion,and so on.To resolve these problems,this paper studies how to detect the object more accurate in the complex environment in real time,and proposes a fast object detection algorithm.The main contents and contributions are as follows.(1)A fast object detection network based on hourglass structure is proposed.It inherits the strategy of single-stage object detection algorithm and directly uses the extracted image features to predict the location and category of the object.The whole model adopts full convolution structure,which insure the detection speed.Aim at the problem of the low accuracy in single-stage object detection algorithms for small-size objects,the hourglass structure,which have great influences on key point detection direction in the model,is used.The measure could generate multi-scale features with advanced semantic information and enable to detect small-size objects more effectively.(2)In view of the inaccurate detection of occluded objects,a multi-task learning based object detection model is proposed.The model utilizes semantic segmentation as an auxiliary task of object detection.By performing the optimization simultaneously of two tasks,the model can extract more detailed features and applicate semantic segmentation at the pixel level.Then,the more accurate detection for occluded objects will achieve.(3)In order to solve the problem of the difficult detection in complex environment,we proposed a new object detection approach based on the integrated attention mechanism.The attention mechanism layer is added to the residual structure,that make the model pays more attention to the object,and make the attention mechanism become a part of the identity mapping,which reduces the impact of image background without causing model degradation and improves the accuracy of the model in the same time.