Multi-object Fault-tolerant Tracking And Trajectory Prediction For Unmanned Vehicles

For unmanned vehicles,it’s important to capture information about the static environment and movement obstacles,especially the latter.And the primary concern is pedestrians and other vehicles.Based on the image sequence acquired by the camera,the historical trajectory of each target can be acquired by multi-object tracking.Combined with other sensors,we can get the trajectory of the object in the real world.Based on the historical trajectory we can make reasonable trajectory predictions,used for the development of obstacle avoidance strategies for unmanned vehicles.There are many problems in multi-object tracking in real world,such as the image captured by the camera is blurred due to the body shake,or the detection algorithm is unstable so that miss some detection.To solve this problems,we add multiple fault-tolerant improvement modules to the original tracking system,which makes it more robust while improving accuracy.In addition,in order to explore the purpose of multi-target tracking,we made predictions on vehicle traj ectories.The main research work and results of this paper are as follows:1)Fault tolerance is applied to the detection failure problem that may exist in the original multi-target tracking system,and the occlusion target tracking algorithm is improved The performance of multi-object tracking system depends largely on the performance of the detection algorithm.We consider fault tolerance for the detection algorithm through the MIL filter,which is used to track the target when the detector fails.For the prediction of occlusion targets,we modify the Kalman filter and use exponential function to fit the rate of change of the object frame.A background compensation strategy based on neighboring target change information is also proposed to eliminate occlusion prediction drift caused by background motion.Experiments were carried out on the MOT and KITTI datasets to demonstrate the effectiveness of the improved scheme.2)A fault-tolerant module is added for the case where the multi-target tracking system may acquire blurred images.An image deblurring algorithm based on generating a confrontation network is proposed.In the real world,unmanned vehicles may cause motion blur on the images captured by the camera due to uneven road surface during driving.We restores blur pictures based on deep learning.Based on the original generated confrontation network,a generation-resistance network based on residual approximation is proposed,and a new loss function is added.The image deblurred data set GOPRO has a good effect.We add deblurring algorithms as pre-fault-tolerant modules in multi-object tracking systems.Experiments on the KITTI dataset show that the blur image fault-tolerant moduleproved in this paper can add the anti-blur capability of the multi-object tracking system.3)Propose a vehicle motion trajectory prediction algorithm based on attention mechanism.Multi-object tracking obtains the historical motion trajectory at the image level,combining with distance sensors can get the historical motion trajectory in the real world coordinate system.We improve the original Convolutional Social LSTM network.First,horizontal attention mechanism is used to obtain the importance of neighboring vehicles relative to the vehicle,the neighbor informa tion is weighted and summed to obtain global features,and then added with the local features extracted by convolution pooling to obtain complete neighbor features.Second,the vertical attention mechanism is introduced to improve the Encoder-Decoder structure used in trajectory prediction.The historical node information most relevant to it is used at each moment of prediction.Experiment on the US101 and 180 datasets provided by NGSIM show that the proposed algorithm can obtain more accurate vehicle prediction trajectories.