Train Driver Detection In Complex Environment

Video-based train driver operation monitoring is one of the emerging requirements for train safety management and driving operation regularization.The reliable detection of train driver is challenge due to the poor visual quality and complex background in train surveillance videos which limit the performance of automatic train driver monitoring.Existing approaches about the human detection cannot be adopted directly to solve the train driver detection problem,we thus propose a novel approach in this thesis to detect the train drivers on still images efficiently.Firstly,we introduce three types of detectors,i.e.,a complete-person detector,an occluded-person detector,and a truncated-person detector,to handle the driver detection with different occlusion and truncation situations.Secondly,we develop a combined detector which integrates the features of the three detectors to achieve reliable detection in the complex train driving environment.Finally,we introduce the strategies of optimal part-subset and coarse-to-fine to optimize the combined detector,which achieves good performance on the detection accuracy and speed.Although the combined detector proposed above achieves excellent detection performance on still images,it cannot satisfy all the requirements of the train driver detection in the videos,especially when the drivers always keep moving(including tiny moving)in the train cab.In practice,the spatial and temporal information of the video have not been fully investigated and utilized to solve this problem.In this thesis,we propose a novel framework named C-STC to handles the complex situations of the train and detect the train driver from the video in a more reliable way.C-STC firstly utilizes the combined detector to detect the train driver as the initial result.Afterwards,the initial result of each frame is processed via the spatial context constraint which restrains the false detection.Finally,an optimal mechanism of dynamic threshold adjustment based on the temporal context constraint is presented,and achieves accurate detection in the whole video sequences.Experimental results demonstrate the effectiveness of the proposed combined detector on still images.As expected,C-STC integrating the combined detector and spatio-temporal constraints achieves better detecting result in the train surveillance videos and enables the proposed framework to work efficiently in real time.