Simulation Analysis Of Multi-object Tracking Method For Autonomous Driving To Cope With Occlusion Problem

Multi-object tracking(MOT),which involves tracking numerous moving targets at once in a video or picture sequence.It is more stringent for the accuracy of MOT in the field of autonomous driving due to the ongoing advancements in various sensor technology.The current occlusion issue,however,poses a significant challenge to MOT,primarily because the information from multiple dimensions and sensors’ sources is not effectively utilised by the methods currently in use,and more valuable information.The most crucial aspect of the Deep SORT(Deep Learning + Simple Online and Realtime Tracking)framework,which is a traditional MOT method,is the addition of epistemic features,which help to solve the occlusion problem by extracting,matching,and updating features.One of the optimization methods of the Deep SORT framework is to optimise the accuracy of feature extraction to make it more accurate in matching;the other is to associative tracking of the front and back frame apparent features of the target,and determination of its motion intention based on the spatial coordinates and inter-frame relationship of the target.Considering these perspectives,this paper suggests a Deep SORT MOT framework based on Transformer feature fusion.This framework maps two modal data into another high-dimensional representation designed for fusion,allowing downstream tasks like MOT to efficiently utilise the original d ata.The following components of specific research are included.Convolutional neural networks and Point Pillars networks are used in the data preprocessing stage to combine the spatial characteristics of the camera and LIDAR sensors to extract the feature vectors of the detection target.The two feature vectors are then position aligned by truncated cone correlation so that both features can find t he corresponding fused coordinates after position encoding.According to the object feature obtained by various sensors,developing a Transformer-based framework for feature fusion,combining temporal and spatial features to build various attention modules,and we also suggest an Io U(Intersection over Union)attention mechanism to locally sample temporal features in order to lessen the computational effort of the entire model.The loss function of the entire model is calculated by classifier and location regression during the decoding step of the fusion process.Afterwards,the gradient is calculated using back propagation,and the parameters are updated using gradient descent to produce the best fused features.Put the fused features into the Deep SORT-based trajectory fitting o ptimisation algorithm framework developed,use the algorithm to determine whether the target is in an occlusion situation,integrate the occlusion target position of each frame into the corresponding trajectory,put the data from the five fi tting fr ames in to th e po lynomial trajectory fitting algorithm,and then form each tracklet.When the occluded target has entirely vanished,in order to maintain data continuity for the occluded target,the position of the occluded target is anticipated by fitting the trajectory and then updated via adaptive Kalman filter.Lastly,in the MOT16 dataset,the MOTA(Multi-Object Tracking Accuracy),IDF1(Identification F-Score),and IDs(ID Switch)are utilised as metrics for measuring the performance of MOT algorithms to compare with other algorithms and read the data to draw the table.In order to verify the effectiveness of the Transformer feature fusion-based DeepSORT MOT framework in simulated vehicles,developing a hybrid Carla-Autoware simulation platform during the test phase.Even after the occlusion condition,the framework can still track multiple targets by keeping an eye on the laser point cloud map.