Pedestrian Detection And Tracking From Monocular Image Sequences

Pedestrian Detection and Tracking are key techniques in intelligent transportation system. Monocular images are easily obtained with a single camera so that detecting and tracking pedestrians in monocular image sequences are more practical than other imageing system. A successful pedestrian detection and tracking system depends on a good feature for detection as well as a reliable tracking method. Firstly, the appearance of pedestrians has large variations and the background is always complex, increasing the difficulty of detecting a pedestrian. Secondly, it is easy to cause drift problem because of occlusions and inaccurate tracking results. For the above problems, this study intends to desire a powerful feature for detection without considering the depth information, and an effective template update strategy to alleviate the drift problem during tracking.Channel features are one of the most effective features for pedestrian detection due to its effectiveness and efficiency. However, this feature is a low-level feature which is lack of semantic information. One possible solution to improve the pedestrian detector is to combine the channel features with mid-level features or high-level features. Recently, deep learning provides promising results in many computer vision applications, which can learn various features from a large number of unlabeled data. Benefitting from the prowess of deep learning, this paper proposed to extract deep channel features to train a pedestrian detector. For each channel, the method learns a group of filters by sparse filtering, and thus generates various higher level channel features. Combining deep channel features and low-level channel features(i.e. LUV channels, gradient magnitude channel and histogram of gradient channels), the proposed method achieves a significant improvement than the baseline detector. The experimental results on the public benchmark show the effectiveness of our deep channel features.In the process of tracking, an alternate template update strategy is proposed to boost an online tracker by alleviating the drift problem. The goal of this strategy is to develop a robust way of updating an adaptive appearance model. There are several modules in the appearance model, and the alternate template update strategy is the case that only one module is updated with newly added data and the others keep stable. It makes proposed tracker be updated while keeping historical information. When occlusions or inaccurate tracking results happen, they affect only one module in the tracker. Therefore, the proposed method can track the object by other modules. An online Na?ve Bayes classifier with proposed deep channel features is trained to build each module in the proposed tracker, and the candidate with the highest classification score is determined as the tracking result. In order to fuse these modules, this paper desired a criterion with considering the appearance similarity and motion information between two consecutive frames. The experimental results show the effectiveness of the proposed tracker and the alternative template update strategy.