Segmentation by weighted aggregation and perceptual hash for pedestrian detection

• A joint BING–HASP framework for pedestrian detection is proposed.
• Segmentation by weighted aggregation is used to segment the salient regions from complex backgrounds.
• Perception hash coded from the segmented image is suitable for pedestrian detection.
• BING can significantly accelerate the computation process.

Main challenges of pedestrian detection are caused by the intra-class variation of pedestrians in clothing, scales, deformations, occlusions, and backgrounds. The prevalent detection frameworks employ a series of dense sliding windows, which are time-consuming. In this work, we equip the detection framework with another new strategy, and extract the new features, to eliminate the above requirements. Segmentation by weighted aggregation (SWA) provides a probability measure to segment objects from complex backgrounds. Perceptual hash (pHash) has shown its power in similar image retrieval because it is modification-tolerant and scale-invariant. The proposed approach uses binarized normed gradients (BING) to efficiently generate a small set of estimation proposals, and formulates SWA and pHash into a joint descriptor, called HASP, to improve the detection performance significantly. Experimental results both on INRIA dataset and ETH dataset have demonstrated the effectiveness and efficiency of the proposed approach.