SRS Image Recognition System
The SRS image recognition system is a system based on camera sensors, which can detect and recognize the dimensions and location of other traffic participants, traffic signs and information, for automated driving vehicles. The advantage of this system is that its accuracy is higher than that of the general image recognition systems. The system mainly consists of three modules: video acquisition module, image entity recognition module and image calibration and size calculation module.
Video Acquisition Module:
The video acquisition module of SRS is mainly aimed at the acquisition of video by camera. Now the standard version of the video acquisition module will complete an acquisition of 30-50 frames video image in 3 seconds, and transfer the key frame to the image entity recognition module, and finally control the system operation time within an acceptable range for the system. The better the performance of the used camera is, the better the effect can be.
Image Entity Recognition Module:
The image entity recognition module of SRS is also the core part of the system. It is mainly designed to recognize the size of the traffic participants (including its height and diameter, or height and width), the center position of the traffic participants (including radial position), and the traffic sign. For this purpose, we provide two solutions. The first one is based on the traditional image recognition technology: template matching solution. The second one is automated image recognition and calibration solutions based on deep learning algorithms: resnet, vggnet and inception. According to the specific requirements of customers, we can select and implement the optimal solution to complete the requirements.
● Solution Based on the Traditional Image Recognition Technology This solution mainly uses some traditional image recognition technologies including preprocessing, feature extraction, classifier design, and template matching. Since the traffic sign and information (including the lanes) in the automated driving environment are relatively in a fixed size, these objects can be easily detected and positioned by matching the real view with the image templates (pictures) taken previously. The advantage of this solution lies in the fact that the amount of data used is small, and a relatively accurate identification system can be constructed quickly. For automated driving, this solution can provide high quality and high efficiency detection of traffic signs and information (including lanes) in the case where the types of the objects to be detected are fixed and the quantity is small. ● Automated Image Recognition and Calibration Solutions Based on Deep Learning Algorithms: Resnet, Vggnet and Inception The solution lies in the use of the latest deep learning image recognition technology to construct an automated image recognition and calibration solution. The specific method in this solution is based on the existing Resnet (depth residual network). It performs a fine-tuning process on pictures of automated driving scenarios and ultimately enables the existing depth residual network to learn the elements in the automated driving scenario pictures, so that the network can be more effectively applied in automated driving scenarios. The so-called fine-tuning and Transfer Learning are two different concepts. However, limited to the field of CNN training, you can basically regard fine-tuning as a method of Transfer Learning. For example, we have a new data set and let engineers do a picture classification. This data set is about cars. The problem is that compared to the wide variety of car models on the market, there are few categories of cars in data, and there are not many data. The training of CNN from zero is very difficult and it is easy to overfit. In this case, Transfer Learning is a good solution. Transfer Learning is performed by using an ImageNet model that has already been trained. The ImageNet may not be used for automated driving, but because ImageNet can have a lot of tagged training data set, the generalization ability of pre-trained models (such as CaffeNet) can be expanded. Ignoring the middle layers, only by fine-tuning the latter layers, the results can usually be very satisfactory. Simply put, we are fine-tuning and adapting our existing scenarios with a model that others have built using millions of images, so that the complex image features of that millions of images can be applied to this case. This makes the system obtain better accuracy when using fewer pictures. Most importantly, the amount of data required to achieve high performance recognition classifier is relatively small. In this way, the automatic driving scenarios can be efficiently and stably modeled to accurately obtain the desired recognition classifier.
Another advantage of this solution is that it can use the GPU to solve very complex matrix and vector operations. It can still achieve faster response speed and excellent recognition quality when the identified traffic participants are various and different in form. For ordinary machine learning image recognition technology, an important difficulty in image recognition in an automated driving environment is that the reflection of ambient light on highly reflective materials can affect image recognition. Our automated image recognition based on deep learning resnet, vggnet and inception can automatically learn these interferences and solve the problem well. Image Calibration and Size Calculation Module: The function of the image calibration and size calculation module of the SRS makes it possible to measure the size and position of the traffic participants. After the image entity recognition module detects the position of the specific objects, this module performs a series of image processing, including pixel calculation, to determine their final calibrated position and size. By using a second template matching the accuracy of the detection can be ensured. After that, the result will be finally outputted. This entire process ensures a stable detection and measurements with high quality.