Research And Application Of Multi-object Tracking Algorithm For Remote Takeover Scenario Of Autonomous Driving

Autonomous driving is a key technology that will change the way humans travel in the future.However,existing technical bottlenecks and progress in laws and regulations are not enough to achieve the popularity of fully autonomous driving,which still requires manual takeover.In this paper,we use multi-object tracking technology to track the objects in the environment around the controlled vehicle and predict the movement trajectory of obstacles,to help the remote receiver perceive the danger and improve the safety of remote takeover.But the multi-object tracking algorithms still face many challenges in autonomous driving remote takeover scenarios,including tracking of dangerous dynamic objects,transient failure of 3D sensors,and continuous iteration capability.Research is conducted to address the above issues,and the main research of this paper is as follows.(1)For the problems of dangerous dynamic object tracking and transient failure of3 D sensors,this paper designs and implements a multi-object tracking algorithm based on dual trackers and D-S evidence theory.The method fuses 2D and 3D sensor information to track the object to achieve the necessary safety redundancy.The two internal trackers accomplish motion prediction and data association for 3D and 2D information,respectively.We propose a two-level conflict resolution method based on D-S evidence theory,which converts various features extracted from different sensors into mass functions to solve the possible output conflict problem of the two internal trackers.On the open-source public tracking dataset,this paper simulates the scenario of 3D sensor failure by ignoring the output of 3D detectors to simulate the failure due to weather,mutual interference,and other factors.Comparisons are made using the latest evaluation metrics,and the results show that the proposed method outperforms other advanced open-source tracking methods in various scenarios.(2)For the problem of continuous algorithm iteration,this paper designs a data closed-loop method based on an autonomous driving simulator,which is used to simulate realistic self-driving car road tests and can automatically collect and label specific scenario data.Special triggers for data return are designed to ensure that the autonomous driving system can actively detect and return corner case data from the massive test data,thus acquiring high-value training data for iterative detection models with low transmission and storage costs.(3)In this paper,the proposed multi-object tracking algorithm is deployed to an autonomous driving simulator to implement an autonomous driving remote takeover system based on the multi-object tracking algorithm.The remote takeover system implements a collision warning module that combines the communication time delay to calculate the collision warning safety distance for hazard warning.The system also implements a takeover control logic module that combines the time delay,current road traffic conditions,and the state of the car to determine whether the control command can fail and whether the takeover is terminated,so that the controlled car has both the full self-driving driving function and can be taken over remotely by the controller without conflicting with each other.Experimental results on both the public dataset and the produced dataset show that the multi-target tracking method proposed in this paper has a good tracking performance as well as a continuous iterative capability and can be applied to the autonomous driving remote takeover scenario to improve driving safety.