Improving Run Time Efficiency of Semantic Video Event Classification

To bring autonomous vehicles on the road requires modern technology which promises precise sensing of the different parameters and accurately using the collected set of information for the course of action. Sensing of the surrounding parameters includes system understanding signal and lighting systems, identifying hazardous situations, distinguishing different obstacles, and according to activating different applications like blind-spot detection, antilock braking, airbags, tire pressure monitoring, battery level monitoring for electric vehicles, downhill control, cruise controlling, emergency braking and many other applications. To implement the titled architecture a case study of the automatic braking system is implemented using a machine learning approach. Specific identification of the car is done using the Haar-Cascade Algorithm. The module is trained by giving numerous positive and negative images. The large set of the data is stored in the Hierarchical Data Format 5 version of the HDF5 file format. The XML file and HDF5 files are then imported and a new video stream for identification of the car and brake light is fed to the module. The prediction of the module is done in four different classes such as brake applied, brake not applied, parking light, Left or right indicator, and light off state. The proposed module identifies the brake light of the car with 99% accuracy.