On Region Extraction And Detection Algorithms For Video Targets

Target area extraction and detection is a topic of concern in computer vision,and is widely used in military and life frontier fields such as video surveillance,pedestrian recognition,and driverless driving.With the widening of application fields and the development of technology,people have higher requirements for the detection speed and accuracy of video targets.The focus of the research is on extracting candidate boxes that may have targets and screening them.Common lighting,viewing angles,background occlusion,and too small targets pose problems for the extraction and detection of target areas.In view of the above problems,this paper studies and improves the target region extraction and detection algorithm in video:(1)A new target region extraction algorithm HSS algorithm is proposed for region extraction.Before the Selective Search algorithm,the Harris corner detection algorithm is used to extract the corner points in the image,and the important features of the image are preserved.The addition of the Harris corner algorithm greatly reduces the original Selective Search algorithm to extract the pseudo-candidate frame,reduces the time complexity,and removes the redundant candidate frame by the non-maximum suppression algorithm to remove the false target area.Finally retain the best results.Experiments on the datasets of normal targets,infrared small targets,and ordinary small targets prove that the algorithm can reduce the time used for region extraction while reducing the false positive rate and not detecting the target,and speed up the subsequent target detection process.Has an indispensable role.(2)In the aspect of target detection,make full use of the advantages of deep forest(gcForest)algorithm in data representation,and extract the higher dimensional features of input image and video through end-to-end training,thus proposing an improved deep forest algorithm.(gcForest-deep).The deep network structure is added to the multi-granularity scanning stage to form a deep multi-granularity scanning structure,and more dimensional raw image information is extracted to input the subsequent cascaded forest parts.The gcForest-deep algorithm is applied to the small target data sets with two target pixel ambiguities.It can be seen from the experiment that the algorithm has higher accuracy in small target detection and is suitable for scenarios with high accuracy requirements.In this paper,the problem of redundant candidate frames in the target region extraction algorithm is proposed.The HSS algorithm is proposed to reduce the detected candidate regions and false targets,and reduce the time complexity,which greatly improves the quality and efficiency of candidate frame extraction.Aiming at the problem that the deep forest algorithm lacks the relevant original image information in the cascade forest part,the deep multi-granularity scanning layer is added into the structure to extract more features and improve the accuracy of target detection.