Method for detecting and recognizing various types of obstacles based on convolution neural network

Abstract

The invention relates to a method for detecting and recognizing various types of obstacles based on a convolution neural network, and belongs to the technical field of computer vision and target detection. The method converts a vehicle video into picture frames, extracts the ROI and the position information thereof in the original image from the picture frames by using an inter-class variance method and morphological operation, and puts the extracted ROI into an AlexNet network to be classified, at the same time, estimates the states according to the classified location information of the obstacles by using Kalman filtering so as to achieve real-time obstacle detection and recognition. The method extracts the image itself and the characteristics of the images, improves the accuracy of the obstacle detection and recognition, sets the information, such as the attribute, the movement tendency and the like of the obstacle, into the whole detection and recognition system, and has an important effect on safe driving for drivers or intelligent vehicles, and plays a vital role for full-intelligent driving systems in the future.

Description

technical field [0001] The invention belongs to the technical field of computer vision and target detection, and relates to a multi-type obstacle detection and recognition method based on a convolutional neural network. Background technique [0002] Object detection and category recognition are one of the core issues in target detection and computer vision. Detecting objects in the driving process, such as people, cars, utility poles, street signs, guardrails and other obstacle information is very important for the safe driving of artificial driving and smart cars. has a vital role. Distinguish the types of moving or stationary obstacles, calculate the distance between the obstacle and the current driving vehicle, judge the movement trend of the obstacle, and design a high-quality visual driving assistance system, which can provide very valuable assistance for smart car obstacle avoidance information. If the detected obstacle is a high-value and moving obstacle such as a p...

AI Technical Summary

Problems solved by technology

Assume that the detected obstacles are of low value such as guardrails and utility poles, and are in a static state, but if a collision occurs, which will cause great property damage and personal injury to the car or driver, it should be biased towards a constant speed or accelerate away from the obstacle treatment measures

[0003] At present, the method of detecting targets based on convolutional neural networks can only detect static objects, and only detects a single target, such as the patent application number 201310633797. Detection method" only detects pedestrians, but simply distinguishes target objects and non-target objects, does not track the target, and does not label the attributes of the target object, such as the distance between the target and the position of the car and the distance between the target and the target. sports trends

And it adopts a single method of extracting image features, and only extracts the pixel value of the detection area in the image as the feature learned by the convolutional neural network.

However, image features based on pixel values ​​cannot describe the local distribution of colors in the image and the spatial position of each color, that is, they cannot describe the specific category of targets in the detection area in the image, the detection and recognition accuracy is not high, and they are easily affected by the environment. Interference, such as lighting, image resolution, image shooting angle, jitter, etc.

Images