| As a basic tool to ensure the safety of residents’ life and property,surveillance video system has become an important part of the development of harmonious society.For accurate object extraction in the surveillance video,the applications of object detection and semantic segmentation have urgent realistic need and important research significance for the target reidentification and cross-region retrieval.Traditional object detection algorithms have poor performance in the multi scale scenario of surveillance video,and they have defects when detecting small objects in this scenario.Traditional segmentation networks have a pretty performance,but they cannot meet the requirements of paramaters and computations,and they have general performance for multi-scale objects.Since there are many multi-scale objects in surveillance video,the detection network is optimized from the aspects of backbone network,loss function and feature fusion,and the segmentation network is optimized from the aspects of light weighted model and feature map upsampling module.From object detection to segmentation,a complete accurate object extraction system in surveillance video is built.To build this system,the following aspects of work are compeleted in this paper:First,the one-stage detection algorithm based on Efficient Net is proposed after comparing and analyzing the classic feature extraction network,and focal loss is introduced to slove the imbalance of positive and negative samples.To improve the recall of algorithm,the prior box is generated using the clustering algorithm.The experiments show that the algorithm proposed in this paper has a better detection performance and these two optimizing strategies can improve the performance of detection algorithm on multi scale scenarios.To futher improve the precision of algorithm and the detection performance for multi scale object.Based on the data set,the distribution of prior boxes on multiple scales is analyzed experimentally and a multi scale feature fusion based on the attention mechanism is designed,this paper improves and uses the loss function based on Io U to improve the positional accuracy of algorithm.The experimental results show that these three optimizations can both improve the recall and the detection performance of the algorithm when detecting multi scale objects in surveillance video.Finally,the semantic segmentation algorithm is used for accurate extraction.After classical semantic segmentation and instance segmentation algorithms are compared and analyzed,this paper chooses UNet.In order to improve the efficiency of algorithm,the light weight model is applied.And the upsampling module based on pyramid pooling module is proposed.The experimental results show that the efficiency of light weight model is greatly improved,and upsampling module based on multi scale pooling can improves the performance of segmentation algorithm. |
