<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Siddhesh Zanj on Medium]]></title>
        <description><![CDATA[Stories by Siddhesh Zanj on Medium]]></description>
        <link>https://medium.com/@siddheshzanj?source=rss-2b40c4c8cc25------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*9TFCsKsh0XBVjRhD.</url>
            <title>Stories by Siddhesh Zanj on Medium</title>
            <link>https://medium.com/@siddheshzanj?source=rss-2b40c4c8cc25------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 06:27:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@siddheshzanj/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Extended Kalman Filter]]></title>
            <link>https://medium.com/@siddheshzanj/extended-kalman-filter-94fe07fd5c79?source=rss-2b40c4c8cc25------2</link>
            <guid isPermaLink="false">https://medium.com/p/94fe07fd5c79</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[udacity]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[infosys]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Siddhesh Zanj]]></dc:creator>
            <pubDate>Sat, 28 Apr 2018 08:46:22 GMT</pubDate>
            <atom:updated>2018-04-28T08:46:22.044Z</atom:updated>
            <content:encoded><![CDATA[<p>This is about an interesting project I did as a part of Udacity Infosys Connect Self Driving Car Nanodegree program. Currently, I am in term 2 of the program which includes Sensor Fusion(LIDAR &amp; RADAR),Localization and Controllers(PID &amp; MPC). All the projects in this term will be in C++. So lets get our hands dirty in C++. In Extended Kalman Filter(EKF) project, we are provided simulated <a href="http://www.archer-soft.com/en/blog/lidar-vs-radar-comparison-which-system-better-automotive">lidar and radar measurements</a> detecting a bicycle that travels around your vehicle. You will use a Kalman filter, lidar measurements and radar measurements to track the bicycle’s position and velocity.</p><h4>What is Kalman Filter ?</h4><p>It’s an <a href="https://en.wikipedia.org/wiki/Algorithm">iterative estimation algorithm</a> that uses a series of measurements observed over time, containing <a href="https://en.wikipedia.org/wiki/Statistical_noise">statistical noise</a> and other inaccuracies, and produces estimates of unknown variables that tend to be more accurate than those based on a single measurement alone, by estimating a <a href="https://en.wikipedia.org/wiki/Joint_probability_distribution">joint probability distribution</a> basically Gaussian over the variables for each timeframe. Not understood this definition ?</p><p>Well don’t worry. Its basically a series of matrix equations to quickly estimate the true value of the object being measured(in our case position) when the measured values(i.e sensor data) has lots of uncertainty or error. Lets try out with an example, suppose we are measuring the temperature(T) of object using a thermometer(it has a certain amount of uncertainty or error). Please refer below diagram. Initially Kalman filter starts out with random estimate temperature with more uncertainty then it gets sensor data(small circles) it quickly narrows down near the actual temperature and finally over time t it gets more accurate estimate of the temperature.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/787/1*iNhsqgip8CsobeYu8yF3nw.png" /><figcaption>Kalman Filter example</figcaption></figure><h4>Why use Kalman Filter and not a simple average of measurements?</h4><p>Whenever we use sensor data it includes a certain amount of error or uncertainty may be due to climatic /weather conditions or manufacturing error or error in processing. We need to estimate those measurements as accurate as possible.</p><p>When we take average of the measurements , we need all the measurements at once. Also, if we consider previous measurements few of measurements can shift the average value from actual value. For eg. we have measurements as 5 , 6, 4, 4 and the actual value is 4.5. When we take average of these values its 4.75. In this case, we are getting our estimate close to actual value but what if we have one outlier measurement 9 included , then our average becomes 5.5. So our error increases due to one such outlier value.</p><p>In real scenario where we are tracking an moving object from our car or any other vehicle , if our sensors send erroneous data or suppose if it didn’t send any data for a particular time we can’t rely on average of previous values. Kalman filter overcomes this problem of outliers. At a particular time, it only considers one sensor data and compares it with previous values such that it gives more weightage to the previously estimated measurement if it has low error and gives more weightage to the newly taken value from sensor if this has low error. In this way it removes the problem of outlier since it only considers a fraction of newly taken value at each time step.</p><h4>Where it is used?</h4><p>Well Kalman filter is used in predicting most of measurements where we get data from hardware i.e sensors etc and we are not certain how accurate our data is. Typically these are the applications of Kalman filter : autopilot in aeroplanes, robotics, <a href="https://en.wikipedia.org/wiki/Satellite_navigation_system">Satellite navigation systems,</a> Tracking of objects in <a href="https://en.wikipedia.org/wiki/Computer_vision">computer vision,</a> <a href="https://en.wikipedia.org/wiki/3D_modeling">3D modeling,</a> Battery state of charge (SoC) estimation etc.</p><h4>So how does it actually works ?</h4><p>Here is the general process flow of Kalman filter algorithm:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2XhsE376_bgaQFNk_YQG3Q.png" /></figure><p>Kalman filter algorithm can be basically organized into following steps:</p><ol><li>Initialize state &amp; covariance matrices : We initialize the state(position &amp; velocity) of the moving object i.e. bicycle when we receive the first sensor measurements.</li><li>Prediction step: We make a prediction of a state, based on some previous values and model.</li><li>Updation step : The predicted location and the measured location are combined to give an updated location. The Kalman filter will put more weight on either the predicted location or the measured location depending on the uncertainty of each value.</li><li>And the process repeats as the car gets another sensor measurements.</li></ol><p>Here is the full overview of the Kalman Filter process with matrices &amp; equations :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I3xbZ5dw6YOg1FRCPiHHUg.jpeg" /></figure><p>You may be wondering how this matrix equations came from. Well, these are just the prediction and updation steps using our motion model equations which are as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*kg39Fh-eo3rDJ8FoJNzzdw.png" /></figure><p>When represented in Matrix form,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/396/1*pg60_PudYF7qbQFmKyMN3A.png" /><figcaption><strong>Also written as X’ = F’X + U</strong></figcaption></figure><p>where X` is the newly predicted state produced by multiply previous state X with F as the transition matrix and adding matrix U which considers acceleration(a_x and a_y). However for this project we are considering velocity as constant hence we are not including acceleration term. However we are still incorporating some uncertainty as state prediction noise W.</p><p>Kalman Gain(KG) : This is fraction is responsible for giving the Kalman filter where to put more weight on either the predicted location or the measured location depending on the uncertainty of each value. If error is measurement received from sensor is low, KG will be close to 1 meaning it will take larger fraction of new sensor measurement in updation step; on the other hand if the error/uncertainty is more it will be close to 0 meaning it will take only a small fraction of it and incorporate in previous estimation.</p><p>Also, when we are estimating there may be some amount of noise or uncertainty or error , this we are incorporating as Q.</p><p>Remember from the definition that Kalman filter considers this noise matrices in the form of Gaussian distribution with mean and std deviations which is explained below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/350/1*U-cR-vP8pYUmLUDwCPv23A.png" /><figcaption>Gaussian distribution example</figcaption></figure><p>The std deviation represents the uncertainty/noise in our estimate. As we follow a number of prediction &amp; updation steps our std deviation becomes less(a narrow Gaussian) meaning we are less uncertain about our measurements and hence gives accurate results. Thats the magic of Gaussians and thats how Kalman Filter works.</p><h4>Now we got a gist of Kalman Filter, so what’s extended version of it ?</h4><p>In Kalman filter, we are assuming that prediction &amp; updation steps are linear functions but most real world problems involve non linear functions. In most cases, the system is looking into some direction and taking measurement in another direction. This involves angles and sine, cosine functions which are non linear functions which then lead to problems. We receive these non-linear measurements especially from radar as laser sensors represent measurements in cartesian coordinate while the radar sensors represent in polar coordinate. Hence, we have to be able to linearly map from the polar coordinate to the cartesian coordinate to help work on Kalman Filter . In Extended Kalman Filter(EKF), it uses the method called <a href="http://mathworld.wolfram.com/TaylorSeries.html">first order Taylor expansion</a> to obtain linear approximation of the polar coordinate measurements in the update. In this process, a Jacobian matrix is produced, which represents the linear mapping from polar to cartesian coordinate, applied at the update step. Hence, the conversion matrix <em>H </em>becomes the Jacobian matrix <em>Hj</em>, while we will use non-linear state transition and measurement functions for both prediction and update respectively. Rest all is same as Kalman filter.</p><p>I have explained above few of these equations but EKF in itself is a vast topic hence cannot be put in a single post. I recommend to check the following links for more detailed explanation on the Kalman filter especially the video tutorials series which explains it elaborately :</p><ul><li><a href="https://home.wlu.edu/%7Elevys/kalman_tutorial/">The Extended Kalman Filter: An Interactive Tutorial for Non-Experts</a></li><li><a href="https://www.youtube.com/watch?v=CaCcOwJPytQ&amp;list=PLX2gX-ftPVXU3oUFNATxGXY90AULiqnWT">Special Topics — The Kalman Filter (Video Tutorial)</a> by Michel</li></ul><p>Here is the result(video) of my project. In this, Lidar measurements are red circles, radar measurements are blue circles with an arrow pointing in the direction of the observed angle, and estimation markers predicted by our code are green triangles. Our aim was to minimize the RMSE (Root mean squared error) values as low as possible to give more accurate results.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FW1on6bgQpEA%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DW1on6bgQpEA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FW1on6bgQpEA%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fdb4ede0a16ae22aa2fe67aefe6ef6e6/href">https://medium.com/media/fdb4ede0a16ae22aa2fe67aefe6ef6e6/href</a></iframe><p>Also, there is one related topic, the <strong>Unscented Kalman filter</strong> or Sigma point filter which solves the non-linearity problem in Kalman filter by using the concept of sigma points. When the prediction &amp; updation steps are highly non-linear, EKF will give relatively poor performance. The Unscented Kalman filter uses a deterministic sampling technique known as the <a href="https://en.wikipedia.org/wiki/Unscented_transform">unscented transform (UT)</a> to pick a minimal set of sample points (called sigma points) around the mean. The sigma points are then propagated through the non-linear functions, from which a new mean and covariance estimate are then formed.</p><p>Feel free to try this project, here is my github link(my code is in src folder) :</p><ul><li><a href="https://github.com/Sidz204/CarND-Extended-Kalman-Filter-Project.git">Sidz204/CarND-Extended-Kalman-Filter-Project</a></li><li><a href="https://github.com/Sidz204/CarND-Unscented-Kalman-Filter-Project.git">Sidz204/CarND-Unscented-Kalman-Filter-Project</a></li></ul><p>Reach out to me incase you have any issues regarding implementation or setting up environment . My linkedin profile : <a href="http://www.linkedin.com/in/siddhesh-zanj-2b934496">here</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=94fe07fd5c79" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Advanced Lane and Vehicle Detection]]></title>
            <link>https://medium.com/@siddheshzanj/advanced-lane-and-vehicle-detection-7d9d1388cec0?source=rss-2b40c4c8cc25------2</link>
            <guid isPermaLink="false">https://medium.com/p/7d9d1388cec0</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[udacity]]></category>
            <category><![CDATA[vehicle-detection]]></category>
            <dc:creator><![CDATA[Siddhesh Zanj]]></dc:creator>
            <pubDate>Sun, 18 Mar 2018 19:51:00 GMT</pubDate>
            <atom:updated>2018-03-18T19:51:00.719Z</atom:updated>
            <content:encoded><![CDATA[<p>This post is about a mashup of two projects i.e. to identify the lane boundaries and detect vehicles in a video which I did as a part of Udacity Infosys Connect Self driving Car Nanodegree.</p><h4><strong>Project 1: Advance Lane Detection</strong></h4><p>In this, the driving lane lines on each side of the car have been detected and visually highlighted the area between them. This is totally based on Computer Vision techniques done with OpenCV using python.</p><p><strong>Step by Step Implementation:</strong></p><p>We design a software pipeline to apply on each frame(image) of video.</p><p>Step 1 (Camera calibration and image distortion) : When a picture is taken, 3D objects in the real world are transformed into a 2D image. This causes distortion i.e we may see a change in the object size increased or distance changed than the actual size and distance. Every camera has its own distortion coefficients. In this case we calculate those coefficients and apply on our image frame to undistort it. The chessboard image is used for reference.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*0i-lZZfo5Ir0U4tnBOAROw.png" /></figure><p>Step 2: Color transforms, gradients or other methods to create a thresholded binary image(refer github link below for details/code)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/378/1*KqQx9f-QLXehyD5ArRR0BA.png" /><figcaption>Undistorted test image</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/378/1*luk0q3ncwMTq-4fqNM7t0g.png" /><figcaption>binary image after applying threshold</figcaption></figure><p>Step 3: Performing Perspective transform (Bird’s eye view of image) and Identify lane-line pixels and fit their positions with a polynomial equation for future detections</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/378/1*LvIT3vxyQlXSZmOAtfkn8w.png" /><figcaption>how lane lines might look from top</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/378/1*nJ_jGnSaxYdaDioF7m4sqg.png" /><figcaption>fitting pixels in polynomial equation of degree 2</figcaption></figure><p>Step 4 :Calculate the radius of curvature of the lane and the position of the vehicle with respect to center. And final display of the output of the detected lanes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/378/1*oGIGbEzltkXJk4MQar4IEQ.png" /><figcaption>final output</figcaption></figure><h4>Project 2 : Vehicle Detection (Using YOLO architecture)</h4><p>In this, we have to draw bounding boxes around the vehicles detected in a video stream. This can be done in two approaches i.e. CNN(Convolution Neural Network)/Deep Learning based approach and the HOG(Histogram of Oriented Gradients) approach. Here, I am posting about the CNN approach which I implemented using YOLO architecture. You can also find the HOG implementation in my github repository in which I used SVM(Super Vector Machines) classifier to train and predict the images based on HOG features.</p><p>About YOLO(You Only Look Once): It is a state-of-the-art, real-time object detection system. YOLO reframes object detection as a single regression problem, straight from image pixels to bounding box coordinates and class probabilities. A single convolutional network simultaneously predicts multiple bounding boxes and class probabilities for those boxes. YOLO trains on full images and directly optimizes detection performance. For detecting just the cars, here we are using tiny-YOLO v1. The main research paper depicting YOLO performance and implementation can be found on <a href="https://arxiv.org/abs/1506.02640">https://arxiv.org/abs/1506.0264</a>0 by Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi. Also check out <a href="https://github.com/pjreddie/darknet/wiki/YOLO:-Real-Time-Object-Detection">https://github.com/pjreddie/darknet/wiki/YOLO:-Real-Time-Object-Detectio</a>n and <a href="https://pjreddie.com/darknet/yolo/">https://pjreddie.com/darknet/yolo/</a></p><p><strong>Main Insight :</strong></p><p>The main idea is to use the trained weights from YOLO system to train the CNN model which give predictions about the bounding boxes along with probabilities and confidence scores giving us insight whether the box contain an object or not (in our case car). Depending on the threshold we set, only the bounding boxes above that confidence score threshold are retained. If there are more than one bounding box on a particular object than the box with maximum confidence score is retained. Here is an image of YOLO implementation taken from research paper.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/526/1*hIML4NnEJtck8WLuGk-VLA.png" /><figcaption>YOLO detecting dog, bicycle and car</figcaption></figure><p>Similar to project 1 , we implement a software pipeline on each frame to get the final output details of which can be found on my github link provided.</p><p>Here is one of the output images when I implemented YOLO architecture for detecting cars in the video stream:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/601/1*IqQgW3AoIVPbKA5pm7b0oA.png" /><figcaption>image frame of the original video</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/601/1*wcCJZ_tFr7_sd6KeKzITUw.png" /><figcaption>bounding boxes detecting the cars</figcaption></figure><p>I have mashed up the two outputs from the project 1 and project 2 and displayed in one single video as follows:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F6s8R3ExWyVY%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D6s8R3ExWyVY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F6s8R3ExWyVY%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/5f172f4bc98785a432b9dbd4dfc9133f/href">https://medium.com/media/5f172f4bc98785a432b9dbd4dfc9133f/href</a></iframe><p>For more insight/code on implementation of these projects visit my github links:</p><ul><li><a href="https://github.com/Sidz204/CarND-Term1P4-Advance_Lane_Detection">Sidz204/CarND-Term1P4-Advance_Lane_Detection</a></li><li><a href="https://github.com/Sidz204/P5-CarND-term1-Vehicle_Detection/blob/master/YOLO_vehicle_detection.ipynb">Sidz204/P5-CarND-term1-Vehicle_Detection</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7d9d1388cec0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Simulate Human driving behavior in autonomous cars]]></title>
            <link>https://medium.com/@siddheshzanj/simulate-human-driving-behavior-in-autonomous-cars-6932d6c9a371?source=rss-2b40c4c8cc25------2</link>
            <guid isPermaLink="false">https://medium.com/p/6932d6c9a371</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[driverless-cars]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Siddhesh Zanj]]></dc:creator>
            <pubDate>Tue, 06 Mar 2018 15:44:40 GMT</pubDate>
            <atom:updated>2018-03-06T15:44:40.958Z</atom:updated>
            <content:encoded><![CDATA[<p>A demonstration using Deep Learning for behavior cloning</p><p>This is about an interesting project called Behavior Cloning which I did as a part of Udacity’s Infosys-Connect Self Driving Car Nanodegree Program. The job was to train a car in simulator (just like a racing video game). For this we need to put our simulator in manual mode where it collects data like steering angle, speed, throttle etc. as you drive your car. This serves as a training data for our model (python machine learning code). A model is trained, validated and tested using Keras(a Neural Network API written in Python). The model will output a steering angle to an autonomous vehicle. Mainly, in this project, image data and steering angles are used to train a neural network and then this model is used to drive the car autonomously around the track in the simulator.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LI7q7hRLR29lVv3EXLEphg.png" /><figcaption>Simulator used for training in this project</figcaption></figure><p>Camera Setup: The car has 3 cameras in the front i.e. center, left and right all pointing straight. Below are the images taken from each of those cameras.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/1*fi-xpKC9WN-NBnlaaP4s7A.jpeg" /><figcaption>Left</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/1*FswB_6FiDJDZ_lmzBj0IyQ.jpeg" /><figcaption>Center</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/1*TaNHnlfAE9AegfH0bCZOww.jpeg" /><figcaption>Right</figcaption></figure><p><strong>Performance on known track</strong> : Here is the video of the car driving autonomously on the same track on which it is trained. The values on the right console are the steering angles and throttle values predicted by the model.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F7DdB2rKwDPY%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D7DdB2rKwDPY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F7DdB2rKwDPY%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/deebc543b47c781802516bbd3e7b2c09/href">https://medium.com/media/deebc543b47c781802516bbd3e7b2c09/href</a></iframe><p>Performance on unknown track : The real test of our CNN (Convolution Neural Networks)model is to predict on unknown data i.e. in this case track. Here is the video of the car driving autonomously on track 2 on which it is not trained.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FjitVBY2GqT0%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DjitVBY2GqT0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FjitVBY2GqT0%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c32aaee2a21d5c73436267e7cc73256c/href">https://medium.com/media/c32aaee2a21d5c73436267e7cc73256c/href</a></iframe><p>You can see that the car drives slowly taking sharp turns just like a human driver when he is exposed to new difficult road. Also, as he gains experience he can drive almost on any road. Similarly, the car is now trained for only one track, if we feed into more training data car will learn more and predict &amp; drive precisely on all tracks.</p><p>Limitations: This model is not so generalized to run on any tracks. We can use more augmentation techniques so as the car drives autonomously on most other tracks. Also, I have not predicted speed and brake values for training and predicting. You can use similar approach for that.</p><p>For more insight on project documentation and code, it is available on <a href="https://github.com/Sidz204/P3-CarND-Behaivoural-Cloning.git">https://github.com/Sidz204/P3-CarND-Behaivoural-Cloning.git</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6932d6c9a371" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>