Research On Algorithm Of Traffic Target Detection Based On Surveillance Video

It is an important task of intelligent road monitoring system to detect traffic vehicles through monitoring video.Having reviewed the previous researches and efforts on vehicle detection from video surveillance,we found that the current popular object detection methods based on deep learning applied in vehicle detection,habitually treated the video frames as isolated static images,inputting it into neural network to extract features for detection,and ignored the information of vehicle motion hiding between video frames.The traditional inter frame difference method converts the vehicle motion information contained in the video frames from time domain to spatial domain,and based on that,we propose in this paper to use the inter frame difference map as the input of neural network where vehicle motion information fits into the feature maps.The experiment using the YOLOv3 object detection framework shows that the effect of vehicle detection can reach the same level by replacing the original image with the inter frame difference image as the input.which proves the conjecture true.Therefore,inspired by the research in house price prediction which suggested multi input neural network structure is better than those whose inputs are from text data only or from picture data only,in this paper,we proposes to reform the network structure to input both the original frame image and inter frame difference image into the network,and in order to do so,we construct two independent backbones respectively to extract the features from the two inputs.The experiment results show that the precision and recall are improved by 0.99% and 1.17% respectively compared with the baseline where original frame image is the only input and the network is original YOLOv3.It is proved that the new network we reformed can get better vehicle detection results by adding a backbone to extract the features of the inter frame difference map.Though the new network gets better results,two backbones slow it down.In order to alleviate this problem and further improve the detection performance of the network,we continues to use Octave convolution to replace the normal convolution in the network,and reconstructs each module of the neural network structure.The experimental results show that compared with the baseline,the maximum value of F1 increases by 3.56%,corresponding to precision 10%,recall 4.01%,AP 2.31%.It is proved that using Octave convolution to replace the normal convolution in the network can further improve the detection results,speed up the speed of the network,and fit the requirements of real-time detection.