CARL-ODD: A Vision Benchmark Dataset of Asia for On-Road Vehicle Detection and Recognition

Abstract

The advancement of self-driving vehicles' ability to perceive their environment is highly dependent on computer vision-based object detection and recognition. Over the past few decades, many substantial object detection and recognition datasets have been proposed, including the Karlsruhe Institute of Technology and Toyota Technological (KITTI) dataset, the Beijing Institute of Technology (BIT) vehicle dataset, etc. However, these datasets consist of smooth and organised urban road traffic of Western nations, ignoring the congested and disordered traffic of Asian nations (i.e., Pakistan, India, etc.). To overcome the above-mentioned issues, a substantial dataset named Control Automotive and Robotics Lab- Object Detection Dataset (CARL-ODD) has been proposed for self-driving vehicles to navigate safely while having a detailed perception of their surroundings in an Asian environment. CARL-ODD dataset was collected from over eighty cities of Pakistan containing highways, motorways, and congested as well as diverse traffic scenarios of urban, rural, and hilly areas. The use of both mono and stereo-vision-based driving videos in the dataset provided a considerable advantage over the state-of-the-art datasets. The collected data were then labelled, and 20,000 representative images were selected for the perception module. Moreover, to establish a baseline for CARL-ODD, we have also compared the various pre-trained single-stage and multistage object detection and recognition algorithms with the dataset. These pre-trained object detection algorithms, such as YOLO and Faster R-CNN, have been implemented to detect and classify various on-road vehicles present in the Asian environment. This approach can save time and resources compared to training a model from scratch. The major findings as a result of the analysis showed that the CARL-ODD dataset demonstrated strong performance in detecting and recognising surrounding vehicles, while achieving 91% average precision for single-stage detectors and 93% average precision for multistage object detectors. This work outperforms the state-of-the-art datasets in terms of vehicle classes and diverse traffic scenarios.