End-to-end multi-target identification, tracking and prediction method

Abstract

The invention discloses an end-to-end multi-target identification, tracking and prediction method, and belongs to the technical field of Internet of Vehicles and intelligent automobiles. The method comprises the following steps: establishing an end-to-end multi-target identification, tracking and prediction model which comprises a target detector, a target tracking module and a trajectory prediction module; the target detection module uses a multi-target detector based on a central point; the target tracking module adopts a graph-based convolutional neural network to track multiple targets; the trajectory prediction module performs motion trajectory prediction on multiple targets based on a graph network, including prediction of a trajectory destination point, information transmission between intelligent agents and generation of a future trajectory; according to the method, end-to-end multi-target identification, tracking and prediction models are taken as a whole, and simultaneous training is carried out by adopting a joint training framework. The three modules are trained at the same time and promote each other, the final trajectory prediction precision is further improved, multi-target trajectory prediction can be better achieved, and the predicted trajectory is more reasonable.

Description

technical field [0001] The invention relates to the technical fields of Internet of Vehicles and smart cars, in particular to an end-to-end multi-target recognition, tracking and trajectory prediction technology. Background technique [0002] With the emergence and proliferation of autonomous driving technology, multi-target tracking and prediction in road scenes has become an important part of semantic understanding of urban scenes. It plays a vital role in intelligent transportation tasks such as analysis. Part of the existing target tracking and prediction models are based on a single target. In the context of Internet of Vehicles and intelligent driving scenarios, tasks such as path planning and navigation need to accurately grasp the positions and trajectories of all target objects in the current road environment. Therefore, single target tracking And forecasts cannot meet demand. The main task of multi-target recognition is to identify all target objects in the curre...

AI Technical Summary

Problems solved by technology

[0009] Aiming at the problem that current technology has inaccurate prediction results for multi-target recognition, tracking and prediction, and cannot be applied to real road scenarios, the present invention proposes an end-to-end multi-target recognition, tracking and prediction method, which integrates multi-target recognition, The tracking and prediction tasks are regarded as a complete task, and an integrated model is established to train the task, and at the same time, the powerful feature aggregation ability of the graph neural network is used to obtain rich features for tracking and prediction, and the optimal prediction results are obtained. It can be applied to roadside infrastructure or vehicle-mounted intelligent facilities to complete multi-target recognition, tracking and prediction tasks

Images