Image Classification for Autonomous Vehicles

Abstract

Object recognition of autonomous vehicles is the process of detecting objects that obstruct movement and then recognize them using sensors, data structures, and algorithms. Object recognition system is an automobile safety system designed for the safety of the autonomous vehicle and other traffic participants and reduces collision risk. Recognition systems in intelligent vehicles are needed to detect and recognize traffic signs, vehicles, and other objects. Road accidents have long been a significant issue involving loss of life and property. So recent years have seen rapid developments in autonomous and semi-autonomous vehicles. Autonomous vehicles are a comprehensive solution built for safety and comfort on the roads. This solution has many challenges. One of these challenges is to detect and recognize obstacles while navigating. The only way to detect and recognize these objects is by sensing them using a sensors device. Therefore, vision systems are an essential part of this type of vehicle. This thesis proposed a vision-based system for autonomous vehicles to recognize objects and traffic lights on the road. The proposed system contains three phases: image pre-processing, feature extraction, and classification. In the first phase, some image pre-processing techniques are applied to prepare and improve the input images, consisting of three stages: convert color images to grayscale, histogram equalization, and image resize. It turns out that the image pre-processing phase plays a significant role in improving system performance in terms of improving image contrast using histogram equalization and increasing processing speed by convert color images to grayscale in addition to standardizing the size of the images. In the second phase, extraction of the features from images using Principal Component Analysis (PCA). This stage affects the efficiency of the system as the important features (Principal Components) are extracted from the image, which increases the recognition rate and reduces the computational operations during the classification process. In the third phase, the extracted features are fed as input to the proposed Convolutional Neural Network (CNN) X model and machine learning classifiers for object classification and recognition. Machine learning algorithms used in this proposed are Random Forest, Random Tree, and Naive Bayes. The results show that the proposed CNN model achieved a high recognition rate where the classification precision rate reached 100%, and the error rate is 0%. The low number of false alarms and the high precision rate proves that the proposed system performs very well in recognizing the objects. In comparison, the three machine learning algorithms achieved a medium precision rate compared to the proposed CNN model, where the precision of each of them reached 75% in the Random Forest algorithm, 63% in the Naïve Bayes algorithm, and 53% in Random Tree algorithm.