Pedestrian Detection And Tracking In Complex Visual Scenes

Pedestrian detection and tracking technology enjoys wide application prospectsand great research value in the fields of intelligent visual surveillance, behavioralsemantic analysis, intelligent assistive driving, robots control, etc. Despite a greatmany current research achievements on pedestrian detection and tracking, lots ofproblems remain unsolved and require further study because of the complexity of thescene and human body’s inherent characteristics. On the basis of deep understandingof computer vision principle, this paper carries out research on pedestrian detectionand tracking in complex visual scenes, in view of region of interest analysis,pedestrian detection and classification and target tracking technology. The researchcontents and major contributions are listed in the below:Firstly, a region of interest detection algorithm based on visual saliency wasproposed. For the complex scenes, the traditional methods like motion detection andedge segmentation can barely fulfill purposes. This paper proposed a saliency jointmodel comprised of color, depth and movement features to calculate the colorsaliency map, which accordingly improved local-target-detection competence. Thesaliency of movement and depth was added into the final calculation model, whicheffectively reduced the significant value of the background, but highlighted theforeground and moving target. This algorithm is of low computational complexity andcould handle approximate real-time video images.Secondly, a fast hierarchical pedestrian recognition algorithm was designed andimplemented. This algorithm employed a two-stage coarse-to-fine classificationstructure. Based on structureless HOG feature, the coarse level classifier projected themulti-dimensional feature to low dimension through weighted Fisher lineardiscriminant, and then trained cascade classifier with the effective use of GAB. Thefine level pedestrian classifier used multi-part Latent SVM algorithm. Theimplementation of this algorithm was optimized due to integral histogram, imagepyramid optimization and multi-scale feature estimation approaches. The proposedalgorithm has the following characteristics:1. the use of coarse-to-fine hierarchicaldetection structure with high detection rate and low false alarm rate;2. two classifiersintroducing no new features;3. the adoption of series of accelerating optimization method with a relative high detection speed.Furthermore, after studying of target tracking method based on templatematching, a particle filter target tracking algorithm combined with color features andSURF local invariant features was suggested. For unknown sparse distribution featurepoints of SURF, this paper advanced a fast observation-probability calculation model,and calculated the joint observation probability of SURF and color histogram basedon a fusion mode of uncertainly measurement. Meanwhile, for targets’ possibleperspectives and structural changes, a SURF feature-template-set update method wasproposed to avoid the drop of SURF matching numbers and the instability in trackingprocess. The conclusion can be drawn from experiments that this approach has greatrobustness for either sunlight or block and certain adaptability towards targets’appearance changes.Finally, based on the aforementioned research, an automatic pedestrianrecognition and tracking system was designed and implemented. Specific methodswere given respectively in the aspects of module design, thread design andmulti-target tracking. Aided with CUDA programming technology, this systemassigned complex features calculation to GPU, which could ease CPU pressure and atthe same time realized the improvement of operation speed.