Research On Grain Depot Vehicle Tracking Technology Based On Deep Learning

Target tracking is an important branch in the field of computer vision,and it has many applications in real life.In order to achieve better tracking results,deep learning technology has been widely used in tracking targets in recent years.Target tracking can be seen as a similarity measurement problem.The siamese neural network has many branches and is better at measuring similarity.So it usually has a good performance in target tracking.This thesis intends to track the grain depot vehicles from a perspective to obtain the running path of the grain depot vehicles under this perspective.The results obtained in this thesis can be combined with the re-identification technology to get the complete movement path of the grain depot vehicle,so as to realize the overall uninterrupted monitoring of the grain transportation vehicle driving process,so that some non-compliance behaviors such as control quantity will not appear in the warehouse entry and exit stage.In this thesis,firstly,the siamese network is studied.On the basis of the siamese network,the attention mechanism is introduced,and a siamese network target tracking algorithm based on the attention mechanism is proposed.Then,in view of the situation that dust is easily generated during the internal operation of the grain depot,which leads to the unclear video of the grain transport vehicles collected by monitoring video and interferes with the tracking of grain depot vehicles.This thesis improves the tracking algorithm and proposes a siamese network grain depot vehicle tracking algorithm fused with re-detection mechanism,in order to have better performance when tracking vehicles in the grain depot.The work of this thesis mainly consists of two parts:(1)The attention mechanism is introduced on the basis of the siamese-like region generation network,and a siamese network target tracking algorithm based on the attention mechanism is proposed.The spatial attention mechanism is used to introduce feature association effects in different regions of the graph,and the channel attention mechanism is used to adjust the size of the channel weight to enhance the multi-channel expression of features.The attention mechanism makes the network model have a better discriminative ability.A large number of comparative experiments are carried out on the OTB-2015 test benchmark.The experimental results show that the improved tracking algorithm is less prone to drift when tracking moving vehicles,and can realize real-time tracking of moving vehicles.(2)Aiming at the situation that fog and dust easily appear when working in the grain depot,which will interfere with the tracking of grain depot vehicles,the tracking algorithm is improved,and the siamese network grain depot vehicle tracking algorithm fused with re-detection mechanism is proposed.Firstly,the histogram equalization algorithm is used on the input video data of the grain depot vehicle to obtain an image that removes part of the fog and dust.Then,three classifiers are cascaded to form a cascaded detector,which is used to achieve re-detection and integrated it into the tracking algorithm.Finally,each algorithm is tested on the test video sequence of the grain depot vehicle.The experimental results show that the tracking algorithm can track the grain depot vehicle accurately in real time.