Beyond estimating discrete directions of walk: a fuzzy approach.

Direction of walk of a pedestrian is vital information in various applications like visual surveillance, traffic monitoring and control, assisted living systems and automated car assistance system. Existing methods of direction of walk estimation exploit inter-frame and intra-frame features of a pedestrian frame sequence to classify the motion among predefined discrete direction classes. However, in order to achieve a robust method to estimate direction of walk, a strong analogy to justify an evident stationary or motion pattern as potential feature for direction estimation is essential. Discrete results of walk direction are famously used as it can be estimated in less time and are preferable for less precision sensitive scenario. However, the intra-frame feature that yields per-frame orientation is underutilized when the walk directions are subjected to discrete classes and hence there is a stringent need to go beyond discrete levels of direction and comment on specific walk direction angles at the same time maintaining a strong analogy to justify the potential of proposed features for direction of walk estimation. With this motivation, the article proposes a type-1 fuzzy approach over apposite inter-frame as well as intra-frame locomotion feature of pedestrian to yield precise direction of walk in terms of fuzzy directions beyond discrete levels of pedestrian walk directions. The method identifies eight directions as potential membership functions. Identified features are subjected to rule-based table for identification and removal of noisy orientation results, decision on membership function and their membership grade generation. The defuzzified result yields crisp direction of walk estimates beyond discrete levels of direction. The enhanced direction of walk estimation results is evident from qualitative comparison from existing research works and is also supported by the simulation performed over different datasets. The proposed method achieves 98.33%, 98.79%, and 100% Rank-2 balanced accuracy while applied on CASIA Dataset A, B, and NITR Conscious Walk Dataset respectively, which clearly points out its better performance than state of the art. [ABSTRACT FROM AUTHOR]