Abstract:
Finding vehicles in bad weather is one of the toughest problems I have observed in computer vision. This problem heavily affects the ability and safety of an autonomous driving system and advanced driver-assistance programs (ADAS). Conventional methods for vehicle detection do not work in inclement weather, as random weather conditions and poor visibility are present. It is my thesis to apply Convolutional Neural Networks in such a way as to increase the accuracy of detection of vehicles under such challenging environments. In the process of this study, a feasible vehicle detection model with CNNs was developed and the developed model was implemented using TensorFlow. The model was trained and tested on a diversified dataset; in that, images were collected under clear, wet, snowy, and foggy conditions. The training set was prepared with data augmentation techniques, which helped to make the model more robust. To the training set, some of the data augmentation techniques included rotation, width and height shifts, shearing, and zooming. The CNN architecture is built with multiple convolutional layers; each is designed with the ReLU activation and max-pooling. These are followed by fully connected layers, which classify the vehicle to the correct corresponding class. I then built the model with performance measures, including accuracy, precision, recall, and F1-score; I also used these measures in training the model, where the Adam optimizer was used. Based on the results, I can say that the CNN model developed by me will give much improvement in vehicle detection accuracy compared to conventional methods under unfavorable weather conditions. The contribution of this study within the purview of development of technology for autonomous driving and advanced driver assistance systems is important; it provides valuable inputs in making the vehicle detection systems developed dependable yet efficient enough for real-world applications.
