Crowd Counting And Multi-Movement Vehicle Counting

In recent years,with the rapid development of deep learning,video-based intelligent analysis system has become a hot research topic.Complex scenarios,various weather conditions,realistic requirements of online processing,the rapid growth in the amount of data,and the popularization of drones and other devices pose challenges.This thesis focuses on the two visual tasks of crowd counting and multi-movement vehicle counting.For the problem of crowd counting based on uav videos,this thesis improves the network structure of CSRNet,and exploits optical flow extraction,low-light image enhancement,semantic segmentation and other modules to do data processing according to the characteristics of the dataset.In the semantic segmentation module,since dataset does not have segmentation annotations,this thesis presents an effective semi-automatic segmentation annotation method,and the improved HRNetv2 network is used to segment the building area to eliminate interference.The final experimental results on VisDrone-CC2020 dataset demonstrate the effectiveness of the proposed method.Aiming at the problem of background interference in crowd counting of uav videos,this thesis designs and implements an improved scheme,which realize adaptive selection of multi-scale features through multi-scale prediction and weight branch,and finally achieve the same effect improvement as the semantic segmentation module,which is worthy of further exploration.For the problem of multi-movement vehicle counting based on urban traffic surveillance videos,this thesis designs and implements an efficient multi-movement vehicle counting system,which also meets the requirements of Internet of Things devices.This thesis uses dark channel prior haze removal algorithm to preprocess the videos under severe weather conditions for improving the video quality,and designs post-processing schemes such as detection filtering and reclassification for the detection module to cope with the inconsistency between test dataset and pre-training dataset.For online trajectory counting task,this thesis designs a complete set of end-trajectory judgment rules and an efficient two-stage classification strategy.In addition,this thesis optimizes the time-consuming modules in the system.This thesis used this method to participate in the 2021 AI City Challenge and achieved seventh place.