Autonomous vehicles (AVs) depend on onboard sensors such as LiDAR, RADAR, and cameras to perceive their environment. However, local sensing is limited by blind spots and restricted fields of view, which can compromise safety in complex traffic scenarios. Collaborative perception has emerged as a promising solution by allowing connected vehicles to share perception information, thereby extending visibility and enhancing overall situational awareness. This dissertation advances the field through the development of novel graph-based feature aggregation and attention-driven fusion approaches tailored for collaborative perception. The work begins with the design of a communication framework built on ZeroMQ, enabling real-time exchange of visual and meta-data between vehicles. To reduce the bandwidth demands of raw sensory data transmission, a custom compression algorithm is introduced, ensuring efficient utilization of vehicular communication channels. The first contribution is FF-GAT, a graph attention-based model that aggregates semantic features extracted from multiple convolutional neural networks (CNNs). By representing feature maps as graph nodes and linking them via similarity-based edges, FF-GAT adaptively fuses multi-level feature vectors, laying the groundwork for graph attention networks (GATs) in collaborative perception. Building on this, an intermediate collaborative perception framework is proposed where ego-vehicle and neighboring features are combined through attention-guided aggregation. This dynamic approach highlights informative spatial regions, achieving notable improvements in object detection on the V2XSim dataset. CollabGAT further extends this concept by incorporating both intermediate feature and positional data exchanges. Using spatial and channel-level attention within a graph-based structure, it strikes a balance between detection accuracy and communication efficiency, crucial for real-world deployment. To handle challenging conditions such as occlusion, sensor noise, and partial failures, the Graph-based Iterative Feature Fusion (GIFF) framework introduces iterative spatial and channel attention. GIFF enhances detection accuracy and computational efficiency, particularly in environments with degraded sensor reliability. Finally, the Delay-Aware Collaborative Perception framework (DelAwareCol) addresses the practical issue of communication delays and asynchronous transmissions. It integrates intra-agent temporal aggregation, inter-agent spatial alignment, and adaptive fusion mechanisms guided by dynamic contribution estimation. Experiments on simulated and real-world datasets demonstrate its robustness under high-latency and localization uncertainty. Overall, this dissertation delivers a comprehensive framework for collaborative perception in AVs. By combining efficient communication, graph-based feature aggregation, attention-driven fusion, and delay-aware strategies, it advances the development of safer, more reliable, and perceptually capable autonomous driving systems.