Pedestrian detection has always been a long-standing research direction in the field of computer vision. In view of the rich diversity of pedestrians in clothing, posture, and gesture of views, it is still one of the most attractive and challenging problems in computer vision.

In this thesis, we investigate learning methods for pedestrian detection under complex environment, aiming at improving the performances of current pedestrian detection methods in the following three aspects:

First, existing pedestrian detection algorithms mainly use low-level visual information such as image pixel intensities for feature extraction. However, we human beings also rely on high-level semantic prior knowledge to understand the scene. From this view, we propose methods that unify the convolutional features and semantic features, aiming at making use of the semantic information around a detection window to help distinguish difficult samples, thereby reducing the false positive rate.

Second, many detection methods utilize the same feature extraction method for pedestrians of different sizes, for which the small pedestrians tend to have too low feature resolution to be correctly classified. We propose a multi-resolution Convolutional Neural Networks (CNNs) based feature extraction method that is aware of the pedestrian size and can adaptively extract multi-resolution features according to the size of the candidate region. The algorithm enhances the feature representation ability of various sizes, thereby improves the detection performance.

Third, pedestrians in practical scenes are usually obscured by neighboring pedestrians or objects. Normally, a very low confidence score will be generated on these occluded areas, leading to missed detections. To solve this problem, we propose a coupled-network with an occlusion handling sub-network using a deformable model. The other sub-network can adaptively extract different convolutional representation for pedestrians of different sizes. In this way, we can systematically solve the problem of detecting small and occluded pedestrians.