Evolutionary Optimization of Deep Learning Models for Vision based Accident Prevention and Smart Traffic Surveillance Systems

Abstract

The increasing use of automobiles as the primary mode of transportation has made a critical focus on safety measures to address the growing traffic-related problems. This thesis presents an exploration of four objectives, contributing to the advancement of vision based accident prevention and the enhancement of vision based smart traffic surveillance through the use of evolutionary optimization techniques and deep learning models. In the first objective, a Genetic Algorithm (GA) based Neural Architecture Search (NAS) is proposed for constructing a Mask R-CNN based object detection model specif ically designed for vehicle brake light detection. By automating the design process, the approach addresses the limitations of manually designed Deep Neural Network (DNN) architectures, leading to superior performance in detecting brake light status for both two wheeler and four-wheeler vehicles. The second objective aims to expand the scope of the first objective for vehicle brake light detection task. This approach involves a NAS with an expanded search space en compassing the backbone architecture parameters and training parameters. We employ a modified Differential Evolution (DE) algorithm for the search strategy. This algorithm in corporates evaluation correction based selection for mutation and species protection based selection, aiming to identify an optimal DNN model. The experiments on two-wheeler and four-wheeler vehicle datasets demonstrate the effectiveness of the proposed method. Fur ther, cross-dataset evaluation and experiments on real-world traffic videos demonstrate the proposed approach’s generalization capability. The third objective introduces NAS based approach using a DNN model, designed for vehicle re-IDentification (reID) task useful for smart surveillance systems. The Grasshop per Optimization Algorithm (GOA) is employed to search for the optimal DNN model, considering both architecture parameters and hyperparameters related to the reID task. The experiments on two vehicle reID datasets demonstrate the effectiveness of the proposed method in automatically discovering optimal models for vehicle reID task. Finally, the fourth objective addresses driver distraction detection task for accident pre vention. Recognizing the limitations of manually developed DNN architectures for this iii task, we employ NAS with an improved GA to design a one-stage object detection model. The proposed approach explores YOLO backbone architecture parameters and training pa rameters. Experimental results showcase the proposed approach’s superiority compared to existing models on driver distraction detection datasets, emphasizing its efficacy in improv ing driver safety. In summary, this thesis offers a comprehensive exploration of evolutionary optimization techniques applied to deep learning models for object detection and reID tasks, contributing to vision based accident prevention and smart traffic surveillance systems. The integration of NAS and evolutionary algorithms across these works demonstrates their effectiveness in automating the design process and improving the efficiency of deep learning models for various tasks.