Visual perception optimization method for autonomous driving based on feature time series correlation

Abstract

The invention relates to an automatic driving visual perception optimization method based on characteristic time sequence correlation, which comprises the following steps of: improving a deep convolutional neural network detection architecture: adding an output target autocorrelation layer, and outputting a target time sequence consistency evaluation index as selectable network branch output; A target detection training database improvement method; The offline training process improvement method comprises deep convolutional neural network training loss function improvement: adding an autocorrelation loss function, and adding a certain weight coefficient into a total network loss function to participate in trunk feature part training. According to the method, the stability of the output result of the detection algorithm is optimized in the training and reasoning stages; The stability of visual target detection result classification and position regression is effectively improved, so that the accuracy and stability of related target distance and relative motion estimation are improved, a more accurate and effective target perception result is provided for automatic driving application, the overall visual perception algorithm performance is improved, and the requirement of automatic driving of an automobile is met.

Description

technical field [0001] The present invention relates to the technical field of visual perception for automatic driving of automobiles, in particular to a visual perception optimization method for automatic driving based on time series correlation of features. Background technique [0002] Intelligence is one of the important trends in the development of the automotive industry today, and vision systems are increasingly used in the field of automotive active safety. Single and dual current view, rear view and 360-degree surround view systems have become the mainstream perception devices of existing advanced driver assistance systems. Existing visual perception systems of this type can provide structured road information (various types of lane lines, etc.) and specific types of target information (various traffic signs, vehicles, pedestrians, etc.). Based on the above perception output results, the corresponding early warning system and active safety system are derived. The ...

AI Technical Summary

Problems solved by technology

The existing vision-based deep learning target detection algorithm still has the following defects in the timing stability and consistency of the detection results: the detection results of the target frames in adjacent frames are inconsistent (even if there is no visible deviation in the picture illumination); the position regression of the same target The timing of the result is unstable (due to the influence of various factors such as angle, illumination, position, etc.)

The above problems will lead to large fluctuations in vision-based target distance and relative motion measurement, which will affect subsequent related application algorithms and cannot meet the needs of automatic driving applications (especially high-speed conditions)

However, the current mainstream machine learning-based detection algorithm training is based on time-series discrete samples, without considering the impact of time-series correlation on target classification and regression output consistency.

Images