A foggy weather hides all distant objects from our eyes as well as from the camera. Restoration techniques can help you perform TSR during such weather conditions.
A foggy weather hides all distant objects from our eyes as well as from the camera. Rain droplets can also be a major challenge, as we really need additional efforts to get back the edges from water droplets. Smog too can kill almost all the edges from the image. In most of the cases, restoration techniques can help during such weather conditions.
The process of fog or rain removal is a complex task. It is not advisable to include a rain removal or fog removal in the same pipeline of TSR. The user has to switch the mode to "Rain" or "Fog" so that the system can pre-process the image accordingly and then go for a traffic sign recognition. There are algorithms which can identify rain or fog automatically and can switch to appropriate modes.
Light fog or a little rain can be handled by vision algorithms, but in the case of heavy rain or thick fog, it is very difficult to get the objects which are hidden behind the raindrops and fog. It is required to explore other sensors based imaging techniques like thermal imaging, LIDAR imaging to really solve the problem of heavy fog, rain or smoke.
Irrespective of any particular scenario, a traffic sign recognition system necessitates a classifier in order to recognise the sign. Since the traffic sign standards vary among countries, a classifier is required to make the system robust in various geographies.
Recognition is a multi-class problem as there are many signs to be recognised. Algorithm should be extendible as different countries have different set of signs. Features for learning should be robust enough to handle various shapes and size.
The neural network is one of the classic classifier used for the traffic sign recognition in literature. Random forest or SVM with handcrafted features like HOG or LBP are other discussed algorithms in the literature. Quite recently, one of the algorithm getting traction for TSR is the deep convolutional neural network, which can learn discriminative and robust features.
An appropriate selection of feature sets and detection algorithms make an efficient traffic sign detection system. A properly trained classifier with an adequate number of datasets is also important features for an efficient TSR. The perfect blend of a detector and a classifier can recognise the static and non-deformable signs accurately making TSR one of the popular algorithms in vision based ADAS domain.
Jiji Gangadharan: With nearly 4 years of experience in computer vision algorithm development, Jiji is working as a software engineer at PathPartner Technology. Jiji combines her work experience as a research associate and her post-grad education in computer vision and image processing in her work on Advanced Driver Assistance System (ADAS) applications.