Construction and Application Research of the Visual Image Obstacle Type Recognition Model Based on the Computer-Expanded Convolutional Neural Network

Abstract

Due to the development of computer vision technology and image processing technology, obstacle recognition technology has been widely used in military and scientific research fields. However, most of the existing image-based recognition technologies are easily affected by environmental factors, which makes the application scenario of this system more fixed and cannot be applied in complex environments. This paper mainly focuses on the traditional obstacle detection and type recognition method recognition accuracy, reliability and universality is difficult to meet the technical requirements of intelligent vehicles and unmanned vehicles, traditional detection equipment cost is expensive, and other problems. There are many traditional obstacle detection methods, which basically start from the color, edge, and other information of the target object to do detection and recognition research, but their recognition accuracy, reliability, and universality are difficult to meet the technical requirements of intelligent vehicles and unmanned vehicles, and the detection equipment is expensive. The dilated convolutional neural network has the ability to learn autonomously, using the original image as input, without the cumbersome preprocessing process and can extract features of the target object one by one to achieve more accurate recognition. This design will be based on the expanded convolutional neural network, design an obstacle type detection and obstacle recognition application with high recognition accuracy, and good generalization, in which this paper applies the hierarchical structure of the expanded convolutional neural network weight sharing to learn the characteristics of various types of obstacles and extract the global features with characterization significance, combined with the ROI algorithm to achieve real-time obstacle detection and high accuracy type recognition. The ROI algorithm is combined to achieve real-time obstacle detection and high-precision type recognition.

Figure 1

Diagram of the human obstacle avoidance process based on visual perception.

Figure 2

Hierarchical structure of image features.

Figure 3

Weight sharing.

Figure 4

General scheme of obstacle detection and recognition.

Figure 5

Dilated convolutional neural network (DCNN).

Figure 6

Flow chart of obstacle detection.

Figure 7

Flow chart of ROI extraction.

Figure 8

Obstacle detection and recognition under normal light simple traffic condition. (a) Vehicle inspection. (b) Pedestrian detection. (c) Human-vehicle interaction detection.

Figure 9

Obstacle detection and recognition under normal light complex traffic conditions. (a) A wide range of vehicles in the driving environment. (b) Human-vehicle interaction in the driving environment. (c) Human-vehicle interaction in the driving environment.