Object And Activity Detection Methods In Complex Scenes

Object and activity detection are the important research directions in computer vision,which are widely used in intelligent transportation,video surveillance,drone vision etc.At present,there are still many challenges in object and activity detection in complex scenes,such as target occlusion,small target scale,unbalanced training sample categories,complex background,changing scenes,and difficulty in temporal description,etc.,which still need further to be explored and researched by researchers to improve the performance of object and activity detection.This thesis conducts an in-depth exploration of object and activity detection tasks that contain these difficulties.Based on the VisDrone-DET 2021 evaluation task and the analysis of difficulties in VisDrone dataset,an object detection scheme based on UAV images is designed and implemented.In the scheme we present optimal strategies for data preprocessing,FPN level,ROI level and HEAD level,which are pluggable.We apply these pluggable strategies to different twostage detectors,and achieve better detection performance by fusing multimodel results.For the problem of few-shot object detection,we design and implement an object detection scheme by multi-stage training.This solution further balances the performance indicators of the base class and the new class by optimizing the unfreezing module and the number of samples in the finetuning stage.We analyze the inherent difficulties of few-shot object detection and adopt a staged loss optimization to improve the detection performance of the new class.The experiments on Pascal VOC data set prove the effectiveness of our method.Based on the TRECVID 2021 ActEV and ActEV-SRL evaluation tasks,we design and implement an activity detection scheme in surveillance videos.The R(2+1)D network and the Cascade-RCNN network are adopted as activity region detector.We also present effective classification schemes for person-object interaction activity and person-person interaction activity respectively.In the post-processing stage,different strategies are adopted to further improve the performance of the overall scheme.The experiments on VIRAT and MEVA data sets,as well as the evaluation results of TRECVID 2021 ActEV and AcTEV-SRL,which our scheme ranked first,prove the advancement of this scheme.