Automatic driving vehicle-mounted multi-target coupling identification and tracking method under heterogeneous traffic flow

Abstract

The invention relates to an automatic driving vehicle-mounted multi-target coupling identification and tracking method under heterogeneous traffic flow, and the method comprises a step of establishinga vehicle-mounted embedded application program resource scheduler, and comprises the steps: carrying out the model pruning and restoration of a driving scene identification model; establishing a resource allocation framework for supporting the dynamic state of the deep learning model application program; and establishing a deep learning model to run a resource allocation scheduler, flexibly allocating resources to the deep learning model running concurrently through the resource scheduler, and outputting an optimized scheduling scheme. Memory occupation and switching energy consumption of thedeep learning model on the mobile vision device can be reduced, flexible resource allocation and accuracy balance are provided, processing delay is reduced, multi-target recognition efficiency of anautomatic driving automobile is improved, automobile-road cooperation processing is more timely and accurate during automatic driving of the automobile, and the safety of the automatic driving automobile and the application prospect in the traffic field are further improved.

Description

technical field [0001] The invention relates to the fields of computer vision technology and intelligent transportation, and a multi-target coupling identification and tracking method for automatic driving vehicles under heterogeneous traffic flow. Background technique [0002] With the development of deep learning technology, various artificial intelligence applications strongly attract people's attention. Autonomous driving, as an important field for realizing intelligent transportation and establishing a powerful transportation country in the future, has attracted more and more attention, and safety has always been a key consideration of autonomous driving technology. Sex is very important. Visual perception is one of the key technologies in autonomous driving. Cameras and sensors acquire data from the environment outside the vehicle and transmit it to the processor at the same time; the recognition algorithm in the processor can identify driving scene objects such as pe...

AI Technical Summary

Problems solved by technology

[0003] Although deep learning algorithms have a good computing rate on PCs and servers, most of the current embedded mobile devices have far less computing power than servers, so that many deep learning network models with excellent performance cannot be deployed on embedded mobile devices.

However, real-time visual recognition is required in self-driving cars, without cloud support, and the computing resources of embedded mobile devices are limited

The current practice is to compress the deep learning model to reduce resource requirements, but it will reduce the accuracy rate, and the resource budget of the compressed model is fixed, that is, static

Then, on the one hand, the deep learning model attaches great importance to resource requirements. The resources required by the vehicle scene recognition system are dynamically changing during operation. When the applications reach the maximum available resources in parallel, the applications will compete with each other for resources. This results in a lower frame rate for streaming video processing; on the other hand, if the mobile vision system has additional resources at runtime, the compression model cannot utilize the additional resources to recover its reduced accuracy

In mixed traffic scenarios, many targets need to be identified. The cameras and sensors of autonomous vehicles are distributed along the front, rear, left, right, and roof of the vehicle. The streaming video data transmitted by each camera must be processed at the same time, which requires real-time recognition by the mobile vision system. And processing, under the limited computing resources of embedded mobile devices on the vehicle side, it cannot well meet the resource requirements of Sineng’s excellent deep learning network model, resulting in a greatly reduced accuracy of scene recognition for autonomous vehicles and increased processing delays , which affects the accuracy of the target and cannot achieve real-time performance

Images