Research On Algorithm Of Pedestrian Detection And Target Tracking In Videos Captured By In-car Cameras

Pedestrian detection and target tracking technology in videos captured by in-car cameras is an important part of the intelligent driving system.The technology can help the driver to grasp the real-time movement information of pedestrians nearby the vehicle and discover the hazards that the vehicle may cause to pedestrians in the traffic road,so as to guarantee the guarantee of traffic safety.Although the existing pedestrian detection and target tracking algorithms can satisfy real-time requirements,the algorithm has low accuracy and algorithm research cannot be directly used to solve practical problems.At the same time,due to internal and external factors such as significant individual differences,irregular movements,ambiguous environments,and limited computing resources,there is still a large gap in the accuracy of existing algorithms when applied to practical problems.With the support of the key science and technology project(2017H6009)of Fujian Province,this paper focuses on improving the accuracy of pedestrian detection and target tracking on the premise of ensuring the real-time performance of the algorithm,and uses several methods to improve the performance of the algorithm for the taffic scene.The specific research content of this paper is as follows:Firstly,a pedestrian detection algorithm is proposed,which based on deep learning with several optimization strategies.In the traffic environment,this paper proposes a pedestrian detection algorithm based on deep neural network,aiming at the problem of low pedestrian detection accuracy and high detection error rate caused by weak traditional pedestrian features.The algorithm combines the characteristics of in-vehicle video and pedestrians,and improves the pedestrian detection performance of in-vehicle video by selecting four types of optimized pedestrian data,adjusting the structure of the in-depth network,optimizing the network input size,and augmenting pedestrian data.Experiments on the vehicle pedestrian dataset show that the accuracy of the improved depth model pedestrian detection is improved by 13.4%compared with the original model.Compared with the current representative pedestrian detection methods,the missed detection and error detection are greatly reduced and the detection speed reaches 31 frames/second,both real-time and accuracy can meet the application needs of intelligent driving systems.Secondly,a context-aware and scale adaptive kernel correlation filter tracking algorithm is proposed.In this paper,an improved tracking algorithm for kernel correlation filter is proposed to solve the problem that the existing tracking algorithm fail to track target in the complicated environment and when the change of target scale.The algorithm is composed of two independent correlation filters working together to achieve the target position and target scale estimation.In the position filter,a training sample is added by selecting a background block with large interference around the target to enhance the discriminating power of the filter against the target and the background.In the scalar filter,a dynamic scaled pyramid model is used to replace the traditional fixed ratio.The pyramid model can accurately capture the variation of the target scale in successive frames and improve the accuracy of the tracker's scale estimation.Experiments on the 27-segment vehicle video show that the algorithm effectively suppresses the interferences of environments and objects.It achieves more accurate and more robust object tracking than several popular benchmarks and it is even more robust to the complex background and scale change target tracking in the vehicle environment.In summary,this paper aims at the problem of pedestrian detection and low accuracy of target tracking,using multiple optimization strategies and using background information with effective scaling algorithms to improve the accuracy of detection and tracking,respectively.Experiments on vehicle datasets show that this method achieves real-time high-precision pedestrian detection and target tracking,and has the possibility of being applied to intelligent driving systems.