Vehicle detection method and device with scene self-adaption function

Abstract

The invention discloses a vehicle detection method and device with a scene self-adaption function, and the method employs a deep learning model, can automatically adjust model parameters in real time according to the change of a scene, and achieves the scene self-adaption function. The method comprises a video processing module and a model updating module. The video processing module acquires an image frame through video acquisition equipment, pre-processes the image and then loads a vehicle detection model to detect a vehicle in the image, and the loaded vehicle detection model is a dynamic model updated by the model updating module in real time; and the model updating module stores the latest image, transmits features learned by the complex network model on the latest image to the simple model by means of a knowledge distillation technology, realizes updating of the simple model, then carries out precision evaluation on the simple model, selects an optimal model for model deployment, and applies theoptimal model to a video processing model. The method can effectively solve the problems of poor calculation efficiency of a complex model and insufficient generalization ability of a simple model, and significantly improves the speed and precision of vehicle detection.

Description

technical field [0001] The invention belongs to the fields of intelligent transportation and machine vision, and in particular relates to a vehicle detection method and device with a scene self-adaptive function. Background technique [0002] Vehicle detection algorithms can be roughly divided into three categories: moving object detection algorithms based on frame difference or background modeling, traditional machine learning algorithms based on feature extraction and vehicle classification, and end-to-end vehicle detection algorithms based on deep learning. With the continuous improvement of computer computing power and the continuous expansion of data scale, vehicle detection algorithms based on deep learning have obvious advantages over traditional algorithms, with high detection accuracy and strong scene adaptability. [0003] End-to-end object detection based on deep learning can be divided into one-stage and two-stage methods. The Two-stage method follows the tradit...

AI Technical Summary

Problems solved by technology

In this type of deployment scheme, the vehicle detection model is fixed and cannot be updated in real time. Due to the constraints of the model's generalization ability, when the scene changes, the detection effect of the model will be greatly reduced.

Images