A Unified Multi-Task Learning Architecture for Fast and Accurate Pedestrian Detection

We present a unified multi-task learning architecture for fast and accurate pedestrian detection. Different from existing methods which often focus on either new loss function or architecture, we propose an improved convolutional neural network to effectively efficiently interfuse the task of detection semantic segmentation. To achieve this, integrate lightweight segmentation branch Faster R-CNN framework that enables end-to-end hard parameter sharing in order boost performance maintain computational efficiency as follows. Firstly, Semantic Segmentation Feature Module (SS2FM) refines features RPN stage by integrating generated branch. Secondly, Confidence (SS2CM) classification confidence fusing it with confidence. also introduce effective anchor matching point transform alleviate problem feature misalignment heavily occluded pedestrians. The proposed lends itself well more robust diverse scenarios negligible computation overhead. In addition, can high low resolution input images, significantly reduces complexity. Experiment results CityPersons Caltech datasets show our method is fastest among all state-of-the-art while exhibiting competitive performance.