Target detection method
First, changes in target detection method
Second, the problem detection algorithm based on the traditional manual feature
1. Set by conventional methods wherein, on the one hand difficult to design, on the other hand there is often a designed feature various problems, such as suited for a particular condition, i.e. not robust, low efficiency
2. to extract the target frame by way of a sliding window, and the goal box to determine the classification of this strategy, processes in the extraction of the sliding window is very tedious time consuming
Depth learning phase, the neural network can replace the traditional manual design features and complete target detection, but this is only convolution neural network for feature extraction, did not change the search box to extract policy target area Essentially, these methods until Fast RCNN It appears, proposed to replace the original sliding window by RPN network policy, to break the bottleneck of technological progress
Third, the basic flow algorithm
Process a:
given an image and extracts the candidate frame, the candidate frame feature extraction and feature determination classifier, a candidate block NMS Finally combined to give the final target output
Process II:
a given image as an input, feature extraction using regression + a target block to extract the target region, using the same NMS finally a candidate block combined to give the final target output
A conventional process suitable for target detection method based on a depth and target detection learning
process using two target detection method based on the depth study
Fourth, the traditional target detection algorithm
- Viola-Jones (VJ): wherein is mainly used in conjunction with FIG adaboost integral classifier performs tasks such as detection of human
- HOG + SVM: mainly used for pedestrian detection, by extracting HOG features pedestrian candidate target binding region and SVM classifier determines
- DPM: HOG based on a variant of the same characteristics, the difference is the inclusion of some additional strategies to enhance the pinnacle of precision, traditional target detection algorithm
- …
Fifth, deep learning target detection algorithm
- One-stage (YOLO and Series SSD): target detection is performed by the position and location of targets either directly,
- Two-stage (Faster RCNN Series): to recommend the candidate region by using the network RPN
Sixth, the traditional target detection algorithm VS depth learning object detection algorithm
