| As video surveillance devices are widely used in many fields such as public safety,intelligent transportation,and safe cities,the intelligent processing of video surveillance data acquired by such devices becomes more and more important.Pedestrian detection technology is an indispensable key link in the implementation of behavior analysis and other intelligent processing functions in video surveillance systems.However,due to pedestrian variability,the complexity of the environment,and information loss caused by pedestrians' occlusion,target detection Pedestrian detection under various scenarios in the field is still a difficult and important point.Targeting the accuracy of pedestrian detection and recall rate for pedestrians in complex environments,mutual occlusion of pedestrians,and changing attitudes of pedestrians,this paper uses the deep learning network to learn from training samples and obtains strong changes in environment and pedestrian attitude to get a robust model.Using the Faster RCNN network model to construct a detector for pedestrian candidates to adapt to pedestrian variability and environmental disturbances,thereby improving pedestrian detection accuracy.Afterwards,according to the block description structure of the pedestrian features in the deformable component model,a method of self-adaptive parameter adjustment of the component model is proposed and the dynamic weight adjustment strategy is constructed.The weight of each pedestrian's component detector is dynamically adjusted to enhance its detection sensitivity when the pedestrian is occluded.It is used to verify the detection results of the Faster RCNN network model and detect pedestrians that are missed due to occlusion,thereby improving the recall rate of pedestrian detection.The experimental results on the Peking University pedestrian detection dataset show that the proposed algorithm can improve the accuracy of pedestrian detection and improve the recall rate of pedestrian detection. |
