Research On Pedestrian And Vehicle Detection Algorithm In Infrared Video Image

Infrared imaging is an emerging technology,which also contains the extensive prospect of general application.According to the object will radiate infrared rays,therefore it can calculate the temperature difference between object and background to form the infrared image by using the imaging equipment.It has been using widely in some other aspects as well,for instance,industry production,military purposes,and medical researches,etc.Compare to traditional visible light reflection imaging,the advantage of infrared imaging is able to achieve imaging effect in a good condition,even during the night time or other dark environment that does not reach enough level of visible light.With economic prosperity and development on urban roads,private vehicles had a rapid rising,however,there were many traffic accidents caused by insufficient light while evening driving.In fact that the infrared technology gave efficient solutions to this problem.In recent years,in-vehicle infrared devices have been used more and more.How to detect the objects around the vehicle,（especially vehicles and pedestrians）under real-time monitoring during driving,then process the detected data within appropriate time and way will be the main studies of this dissertation.The dissertation focuses on the low recognition rate of vehicles and pedestrians in the infrared imaging procedure and researches the detection algorithm,based on depth study.To modify the algorithm of the depth study YOLOV4-Tiny and YOLOV3 also make an experimental data comparison.In order to have the network model obtain good performance,I select 10,000 infrared images and use mark tools to highlight,building the new data set.To enhance the model test effect,I use a clustering algorithm to re-calculate anchor box point,to enable common maximum suppression change to soft non-maximum suppression,thus improving the detection accuracy when objects overlapped.In lightweight network YOLOV4-Tiny,to enhance the detection accuracy of the small target in-network infrared video images,here need to require network structure transformation,added attention mechanisms enable network gains better detect effects.Using Depth separable convolution instead of ordinary convolution layer,thereby reducing network model parameters and model size.For the YOLOV3 algorithm,which has a larger network model.It requires replacing core network,for example,using Efficientnet replaces the original Dark Net-53,sequentially decrease the size of the network model.Preventing the potential overfitting happens.Multiscale training is adopted during training to improve the robustness of the network enables the network to have a better detection effect on images of different resolutions and guarantees the original real-time performance of the algorithm.The experiment indicates that all parameters of the improved YOLOV4-Tiny algorithm are better than the previous.After replacing the core network,YOLOV3 reduces the model volume by about 3/4 while ensuring accuracy.In the conclusion,the improved version of the algorithm in this thesis contains light size and can be directly applied to in-vehicle equipment,which illustrates high practical value.