Publication

Abstract : Ensuring pedestrian safety is a critical prerequisite for the widespread deployment of autonomous vehicles, particularly in densely populated and often unstructured traffic environments characteristic of urban India. A significant challenge for current advanced driver assistance systems (ADAS) and autonomous driving technologies is the lack of robust and efficient models capable of accurately predicting pedestrian crossing intentions within these highly dynamic and diverse urban scenarios. Existing prediction models, primarily developed and evaluated on data from more structured environments, frequently exhibit reduced performance when faced with the complex and unpredictable pedestrian behaviors prevalent in such challenging conditions, such as those in India, highlighting a substantial research gap. To address this critical need, we propose a novel hybrid framework for pedestrian intention prediction specifically designed for complex urban Indian traffic. Our method integrates essential visual and behavioral cues by combining YOLO (you only look once)-based pedestrian detection with MediaPipe for detailed pose estimation. Features extracted from these perception modules are then processed by a custom-designed neural network, trained to jointly analyze pedestrian behavior and predict their intention to cross the road. A comprehensive dataset is developed comprising approximately 240 video clips captured from authentic Indian urban traffic environments, offering a unique resource that accurately reflects the diversity and complexity of real-world scenarios. On our custom Indian dataset, the proposed intention prediction neural network (IPNN) achieves an accuracy of 90.72%. This represents a significant advancement over the average accuracy of other state-of-the-art models evaluated on the same dataset (average accuracy: 69.09%), demonstrating an improvement of approximately 21.63%. Furthermore, when evaluated on the public joint attention in autonomous driving (JAAD) dataset, IPNN achieves 90.76% accuracy, demonstrating an improvement of approximately –16.24% over the average accuracy of other state-of-the-art models evaluated on JAAD (average accuracy: 74.52%).