Research On Methods For Pedestrian Detection In Complex Visual Environment Based On Deformable Part-based Models

Pedestrian Detection has applicated widely in human motion analysis, intelligent video surveillance, intelligent transportation and automotive driver assistance(Driverless vehicles).It has great academic value and broad market prospects. Pedestrian detection in complex visual environments will face greater challenge due to complex scenes, illumination changes, scale changes, viewpoint changes, pose changes and pedestrian occlusion. Occlusion is the most common and the most difficult problem in pedestrian detection. Unfortunately, there yet no common solutions for this problem in current domestic and international researchs. In this paper, we research on methods for pedestrian detection in visual-occluded environment.The thesis contains three main work:(1) We proposed DPSM model for occluded pedestrian detection. Compared with the DPM model, it can deal with a variety of pedestrian occlusions better and faster. Meanwhile DPSM method has two advantages: First, we do not need to re-train the model. We reconstructed multiple templates for different pedestrian occlusions based on the existing model. Second, the partfilters in different templates are highly shared. So the number of parts does not change with increasing of the number templates. As a result, it greatly speed up the sliding window detection. Experiments show that, my DPSM model has achieved significant improvement in efficiency and effectiveness of pedestrian detection;(2) We proposed a new MOP strategy in NMS operating to select the best results from a number of classification positive examples. Compared with overlap rate of windows, MOP strategies can select the best window and remove the overlapping windows. On the post-processing of detection results, we modefied the detection results according to the position of the partfilters. We evaluated the MOP strategies and detection window correction, experiments suggested that our methods are simple and effective;(3) We proposed fast DPSM pedestrian detection methods by combining rapid FHOG feature pyramid and cascade detection framework. We use two acceleration mechanism for accelerating our DPSM pedestrian algorithms. In the feature extraction stage, we use a fast HOG pyramid method that estimates the entire feature pyramid through a few layers. experiments shows that it speedups 5 times. In the sliding window detection stage, we adopted the star model cascade detection framework to proposed a cascade DPSM method. The cascade DPSM speedups 19 times in detection stage. Overall, our fast DPSM pedestrian detection method accelerated 14 times in all.