Autonomous Vehicle Along with Surrounding Agent's Path Prediction Using CNN-Based Models

In the modern era, Autonomous Vehicles (AVs) are pivotal within the transportation system, presenting notable engineering challenges in the development of self-driving systems, particularly in agent detection such as pedestrians, surrounding vehicles, traffic lights, bicycles, etc. The dataset utilized comprises level 5 data, encompassing three key tables: scenes, time frames, and traffic agents. Specifically, the traffic agent table contains Zarr compressed files providing information about aerial maps, velocities, and coordinates of agents. The primary objective of this paper revolves around constructing a Convolution Neural Network (CNN) model to discern the motion trajectories of agents in relation to the AV, a demanding and high-priority task in building an advanced self-driving system. To predict these motion trajectories, baseline CNN models like ResNet34 and ResNet50 were employed, compared to determine the optimal model. The results indicate a significant enhancement in Negative Multi Log Likelihood when compared with non-frozen ResNet models. Our future research aims to establish a pipeline for object classification around AV and implement Generative Adversarial Network (GAN) on noise data to project future trajectories.