Research On Multi-object Tracking Algorithm For Intelligent Driving Based On Long-short Term Memory Network

The intelligent development of the vehicle has become the trend of this industry.Environmental perception is an important step in the operation of intelligent driving vehicles,and also the basis for the normal operation of other modules of the vehicle.Its development level directly determines the working conditions of intelligent driving vehicles.During these years,deep learning based methods have achieved a lot in the field of computer vision.The visual perception technology based on deep learning is currently a key technology in those vehicles.Stable and fast tracking of targets in the surrounding environment of the vehicle can provide it with enough information for trajectory prediction.This paper mainly focuses on the object detection and multi-object tracking algorithms which based on deep learning,and tries to build a multi-object tracking model to solve the research difficulties in these fields.For the detection-based multi-object tracking algorithm,an object detection algorithm with good performance is first required as the basis.The research in this paper is based on the Center Net algorithm,starting from reducing the computational cost of the algorithm,and to optimize the design of the model.By replacing the residual module and activation function in the original network structure,the calculation amount of the algorithm is reduced without affecting the accuracy of the algorithm.By enumerating and comparing the data sets that under the intelligent driving scenarios,we use the processed COCO 2017 and KITTI data sets to train the object detection algorithm.The results verified that the improved algorithm has higher accuracy and can correctly identify targets in complex scenes compared with the previous algorithms.For the research based on the recurrent neural network algorithm,this paper constructs a multi-object tracking framework which based on the long and short-term memory network.The algorithm uses the encoder-decoder structure to extract the motion characteristics of the object.Then calculates the similarity score according to the appearance characteristics and the motion characteristics to complete the data association and matching tasks between targets.In the process of feature extraction,an attention mechanism is designed to further learn the correlation between time series information and change the weight of different feature vectors in the tracking process,so that the features can be better used in the tracking process.In view of the frequent change of the target ID value of the previous algorithm,a counter is set in the process of data association to reasonably distinguish between falsely detected,missed and occluded targets,thus to improve the tracking performance of the algorithm.Finally,this paper conducted a number of experiments on the tracking algorithm.Based on the characteristics of the tracking tasks and research method,we selected the evaluation index of the tracking algorithm reasonably.The tracking effects of the algorithm was demonstrated in the process of comparison with other algorithms.At last,part of the tracking results of the algorithm are displayed to verify its great performance.The algorithm shows a good effect on the progress of distinguishing multi-scale targets after training on different data sets.Under multiple scenarios,such as partial occlusion of the target,it can still keep the tracking of the target unchanged,which proves its great tracking effect.