Detection and pose estimation of auto-rickshaws from traffic images

In intelligent transport systems, detection and identification of vehicle types enact a substantial role. In this context, this paper addresses the detection and pose classification of a specific vehicle type: auto-rickshaws which have been heavily neglected by the publicly available vehicle datasets, but remains the most commonly used and cheap form of transportation in south Asian countries. Here, we introduce a dataset for auto-rickshaws which consists of instances of varying shape, orientation, size, scale, colour, viewpoint and many more. Further, we carry out a detailed analysis on the performance of state-of-the-art detection algorithms based on both hand-designed and deep features on the proposed dataset. The introduction of pose classification along with the detection eventually results in better understanding of road scenes involving auto-rickshaws. As a matter of fact, we came up with revisions for the currently employed detection algorithms to achieve a low miss rate on the validation sets. It is evident that the findings of this study are tangible and enormously consequential to the road scene understanding and intelligent transportation of developing countries where auto-rickshaws play a pivotal role in public transportation.