The Research of Vehicle Tracking Based on Difference Screening and Coordinate Mapping - Advances in Image and Graphics Technologies - page 317

In this essay, 11211 samples were collected by a vehicle traveling data recorder,
including 3037 positive samples (figure 3) and 8138 positive (figure 4) samples, and
then we have trained these samples and obtained a vehicle identification classifier.
Fig. 3. Positive Sample Examples
Fig. 4. Negative Sample Examples
In theory, using haar features adaboost classifier to identify vehicle can achieve
high recognition rate, but in the final analysis, the adaboost classifier method for the
vehicle identification is achieved based on a series of fixed threshold of haar features.
If some areas in the image meet the corresponding threshold requirements of haar
features, those areas are identified as vehicles, but the haar eigenvalues of some
non-vehicle area in the image can possiblely meet the corresponding threshold
requirements, thus, in addition to the correctly identified vehicles, there may be some
non-vehicle areas are mistakenly identified as the vehicle, as the black circles shown
in figure 5, but the error detection is likely random; in contrast, the correctly identi-
fied goals change successively between different video frames.
Fig. 5. Result of Vehicle Recognize