Overview / Usage

Integrating Computer Vision and Deep Learning has the potential to bring about a relatively inexpensive, robust solution to autonomous driving. Neural Networks requires large data sets to represent all possible driving scenarios. CNN can be used to perform lane and vehicle detection while running at frame rates required for a real-time system.

In this project highway roads used for detection, highways tend to be more predictable and orderly, with road surfaces typically well maintained and lanes well-marked.

Methodology / Approach

​

Methodology: (CNNs used for detection and recognition)

Vehicle Detection:

​

Convolutional Neural Networks (CNNs) have had the largest success in image recognition in the recent years. From these image recognition systems, a number of detection networks were adapted, leading to further advances in image detection.

Using Overfeat CNN detector, which is very Scalable, and simulates a sliding window detector in a single forward pass in the network using convolution results. There are other detection systems , such as R-CNN, rely on selecting as many as candidate windows.

Overfeat converts an image recognition CNN into a "sliding window" detector by providing a larger resolution image and transforming the fully connected layers into convolutional layers.

CNN takes an image as input and outputs an object mask through regression, highlighting the object location. See the following image of mask detector in gallery.

​

Following is the flow for vehicle detection in terms of sliding window mask detector and bounding box regression.

"Sliding Window" -> "Mask Detector" -> "Bounding Box Regression"

You can see the images of sliding window mask detector and bounding box regression.

​

Lane Detection:

Whereas regression for the vehicle class predicts five dimension value ( 4 for bounding box and one for depth), the lane regression for the vehicle class six dimensions, of which first 4- dimensions indicate the 2 end points of a local line segment of the land boundary, 2 - dimensions indicate the depth of the endpoints with respect to the cameras.

Techniques involved for regression:

Similar to vehicle detection, we use L1 loss to train the regressor.

Gradient Descent : mini-batch stochastic gradient descent for optimisation.

Learning rate : controlled by a variant of the momentum scheduler

DBSCAN to cluster the line segments into lanes.

​

Technologies Used

Framework(s) / topics or concepts/ systems

Convolutional Neural Networks

Overfeat CNN Detector,

Object Detectors,

bounding box regression,

DBSCAN

Keras,

Tensorflow,

Carla : High Definition Simulator

​