Deep Learning for Autonomous Vehicle Surrounding Object Classification and Tracking

Abstract

The safe operation of autonomous vehicles (AVs) relies on accurately interpreting the surrounding environment, including the classification, detection, and tracking of other moving objects such as pedestrians and vehicles. The capacity of an AV to distinguish and track these other objects in real time is critical in making decisions that lead to safe maneuvering strategies in complex mixed-traffic scenarios. If an AV is untrusted to make these decisions in dynamic scenarios, the trust of passengers and pedestrians, as well as of other road users, will be limited. One way to solve this problem is through the use of the deep-learning-based approaches.[2] While the use of deep learning (DL) as opposed to traditional algorithms has led to considerable improvements in different tasks in the recent past, it is still limited by challenges such as the requirement of large, balanced datasets, extensive parameter optimization, and the potential overfitting of the trained model to the specific scenario in which it has been trained. DL might not work well when different data distributions are encountered, and in case of lack of data in different categories, there is a high degree of bias towards particular object classes as seen in.Existing long-term target tracking methods face critical challenges, such as bounding-box drift, occlusions, and the need to establish reliable correspondences between epochs. The vehicle tracking modulates lane assignment by bridging the duplicate vehicle detections across consecutive images. The vehicle classification model assigns labels and is trained for small datasets with reduced biases towards classes with a higher density in the dataset. The resultant system demonstrates a highly interactive vehicle tracking approach, with the augmented classification helping the end-to-end capabilities in dense traffic scenarios.