The project focuses on object detection and classification along with drivable road detection, which is one of the core foundations for autonomous vehicles.
The system consists of 2 neural network pipelines -
The first neural network detects and classifies various objects (traffic lights, pedestrians, cars, trucks, bus) using a single end to end neural network. The neural network is constructed using depth wise and point wise convolutions to increase the speed and reduce the number of parameters. The training of the neural network is carried on google cloud, using ADAM optimization and KITTI dataset.
The second neural network detects drivable lane paths using semantic segmentation based neural network. This network is also trained on google cloud on a custom dataset which consists of videos captured under different environmental and driving conditions.
Overall the first version of the system shows impressive results which could further be boosted. for object detection and classification, improved results can be obtained using tracking algorithms whereas for lane detection larger and more broader dataset can be used for improving the results, apart from changing the structure of the network.